<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sreejith Gopinath on Medium]]></title>
        <description><![CDATA[Stories by Sreejith Gopinath on Medium]]></description>
        <link>https://medium.com/@sreejith.gopinath?source=rss-97dd7846468b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*i8sqBcAi72dyo8tq</url>
            <title>Stories by Sreejith Gopinath on Medium</title>
            <link>https://medium.com/@sreejith.gopinath?source=rss-97dd7846468b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 23:48:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sreejith.gopinath/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Password Was “password1234”. It Was Shipped That Way on Purpose.]]></title>
            <link>https://medium.com/@sreejith.gopinath/the-password-was-password1234-it-was-shipped-that-way-on-purpose-caef89507bec?source=rss-97dd7846468b------2</link>
            <guid isPermaLink="false">https://medium.com/p/caef89507bec</guid>
            <dc:creator><![CDATA[Sreejith Gopinath]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 15:30:20 GMT</pubDate>
            <atom:updated>2026-06-01T15:30:20.603Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A static analysis of MLflow found nine security findings. What happened when I reported them is the more interesting story.</em></p><p><strong>Author:</strong> Sreejith Gopinath<br> <strong>Published:</strong> May 27, 2026<br> <strong>Methodology:</strong> Static source code analysis<br> <strong>Disclosure:</strong> Filed via GitHub Security Advisories, April 27–30, 2026. Vendor engaged May 8, 2026. All findings resolved same day.</p><h3>Background</h3><p>MLflow is an open-source platform for managing the machine learning lifecycle, originally created by Databricks and donated to the Linux Foundation in 2020. It is widely deployed in data science teams, research institutions, and production ML infrastructure, including systems that store credentials, model artifacts, and API keys for connected LLM providers.</p><p>In late April 2026 I conducted a static source code analysis of the MLflow codebase and identified nine security findings spanning hardcoded credentials, OS command injection, code injection, missing authorization, and cryptographic misconfiguration. I filed all nine via GitHub Security Advisories on April 27–30, 2026.</p><p>For ten days, nothing happened.</p><p>Then one escalation email to Databricks security changed everything.</p><h3>The Disclosure Timeline</h3><p><strong>April 27–30, 2026:</strong> Nine GitHub Security Advisories filed against mlflow/mlflow. Maintainers auto-added as collaborators. No engagement.</p><p><strong>May 4, 2026:</strong> GitHub Security Lab confirmed they cannot issue CVEs without maintainer involvement and directed me to MITRE. Nine CVE IDs requested from MITRE (case MCID15746224).</p><p><strong>May 8, 2026, 10:27am:</strong> One email sent to bugbounty@databricks.com referencing all nine advisories and the MITRE submission. AK(pseudonym) at Databricks responded within the hour, flagging the advisories to the maintainer team.</p><p><strong>May 8, 2026, 8:15pm:</strong> B-Step62 (YW, Databricks) began reviewing advisories.</p><p><strong>May 8, 2026, 8:41pm:</strong> All nine advisories closed. Five pull requests opened. All co-authored with or credited to @sreelim.</p><p>Twenty-six minutes from first response to complete resolution.</p><h3>The Findings</h3><h3>Finding 1 — Critical: Hardcoded Default Admin Credentials</h3><p><strong>GHSA:</strong> GHSA-7v72-hrvh-8fjh | <strong>CWE:</strong> CWE-798 | <strong>CVSS 3.1:</strong> 9.8</p><h4>Description</h4><p>MLflow ships mlflow/server/auth/basic_auth.ini with hardcoded administrator credentials publicly visible in the source repository:</p><pre>[mlflow]<br>admin_username = admin<br>admin_password = password1234</pre><p>At startup, mlflow/server/auth/__init__.py line 3919 calls create_admin_user(auth_config.admin_username, auth_config.admin_password) verbatim with no check for whether the default password remains unchanged and no warning emitted.</p><h4>Reproduction</h4><pre>mlflow server --app-name basic-auth --host 0.0.0.0 --port 5000</pre><pre>curl -u admin:password1234 <a href="http://localhost:5000/api/2.0/mlflow/experiments/list">http://localhost:5000/api/2.0/mlflow/experiments/list</a></pre><pre>curl -u admin:password1234 \<br>  -X POST <a href="http://localhost:5000/api/2.0/mlflow/users/create">http://localhost:5000/api/2.0/mlflow/users/create</a> \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{&quot;username&quot;: &quot;attacker&quot;, &quot;password&quot;: &quot;owned&quot;, &quot;is_admin&quot;: true}&#39;</pre><h4>Vendor Response</h4><p>“We don’t consider this a remote vulnerability. To be exposed, an operator must enable basic auth and skip the documented admin_password override. The CVSS 9.8 with PR:N doesn’t apply here, since the principal is the deploying admin, not an external attacker. That said, the warning gap is a fair point. We’ll land a startup check that warns when the default password is in use.”</p><h4>Assessment</h4><p>The PR:N argument conflates operator prerequisites with attacker prerequisites. Once basic auth is enabled with default credentials, exploitation requires no attacker privileges — that is definitionally PR:N. The operator’s configuration mistake is a separate dimension from the attacker’s required privilege level.</p><p>PR #23182 adds a startup warning that fires when admin_password == &quot;password1234&quot; and is marked as a bug fix for the next release. The fix is the right remediation even if we disagree on CVE classification. Notably, the commit is co-authored by @sreelim.</p><p><strong>PR #23182:</strong> <a href="https://github.com/mlflow/mlflow/pull/23182">https://github.com/mlflow/mlflow/pull/23182</a> <strong>Status:</strong> Merged in MLflow 3.12.0. Co-authored by @sreelim.</p><h3>Finding 2 — High: Shell Injection via -P Parameter Keys</h3><p><strong>GHSA:</strong> GHSA-w8c3-p3fc-fv5g | <strong>CWE:</strong> CWE-78</p><h4>Description</h4><p>MLflow’s mlflow run -P flag interpolates parameter key names unescaped into a bash -c invocation. _sanitize_param_dict in mlflow/projects/_project_spec.py:305 quotes values but not keys. compute_command then joins these into a shell string passed to bash -c.</p><h4>Vendor Response</h4><p>“The principal supplying -P key=value is the user invoking mlflow run from their own shell — they already control the host process. There’s no privilege boundary crossed. The quoting fix is trivial and worth landing as code quality. We’ll apply shlex.quote to keys in a regular PR. Crediting you in the commit.”</p><h4>Assessment</h4><p>The self-injection argument is correct for direct invocation. The finding has more force in CI/CD pipelines where parameter keys come from upstream job configurations, but I accept the framing for the primary use case.</p><p>PR #23181 landed the fix. GitHub Copilot’s review noted the tests used shlex.quote (POSIX) where MLflow&#39;s cross-platform mlflow.utils.string_utils.quote should be used for Windows compatibility -- a gap worth addressing before merge.</p><p><strong>PR #23181:</strong> <a href="https://github.com/mlflow/mlflow/pull/23181">https://github.com/mlflow/mlflow/pull/23181</a> <strong>Status:</strong> Merged in MLflow 3.12.0. Co-authored by @sreelim.</p><h3>Finding 3 — High: R Code Injection via Unsafely Interpolated Host</h3><p><strong>GHSA:</strong> GHSA-pv9m-v5wv-j2xw | <strong>CWE:</strong> CWE-78</p><h4>Description</h4><p>mlflow/rfunc/backend.py interpolates the host parameter directly into an R expression passed to Rscript -e:</p><pre>command = &quot;mlflow::mlflow_rfunc_serve(&#39;{}&#39;, port = {}, host = &#39;{}&#39;)&quot;.format(<br>    quote(model_path), port, host<br>)</pre><p>A value like &#39;); system(&quot;id&quot;); cat(&#39; closes the R string literal and injects an arbitrary R statement.</p><h4>Vendor Response</h4><p>“The injection works syntactically, but exploitation requires the first R expression to return — which doesn’t happen in mlflow_rfunc_serve() because it enters a blocking server loop. In the default production path this doesn’t fire. The host parameter also comes from a local CLI flag, supplied by the user running their own server — same self-injection pattern.”</p><h4>Assessment</h4><p>The blocking server loop detail is a legitimate technical constraint I missed. Credit to the vendor for identifying it. The self-injection argument also applies here.</p><p>PR #23179 introduced _r_quote() -- a correct R string literal escaper rather than a naive shlex.quote approach. Copilot flagged that _str_optional() in the same file still uses wrong quoting context and may allow injection in predict(), a gap that should be addressed in follow-up.</p><p><strong>PR #23179:</strong> <a href="https://github.com/mlflow/mlflow/pull/23179">https://github.com/mlflow/mlflow/pull/23179</a> <strong>Status:</strong> Merged in MLflow 3.12.0. Co-authored by @sreelim.</p><h3>Finding 4 — High: Shell Injection in PyFuncBackend via Unquoted local_path</h3><p><strong>GHSA:</strong> GHSA-cw8f-5wj5-grm3 | <strong>CWE:</strong> CWE-78</p><h4>Description</h4><p>PyFuncBackend.serve_stdin in mlflow/pyfunc/backend.py:351 passes local_path unquoted into a shell command string:</p><pre>command=f&quot;python {_STDIN_SERVER_SCRIPT} --model-uri {local_path}&quot;</pre><p>The string is passed to Environment.execute which invokes it via bash -c.</p><h4>Vendor Response</h4><p>“The threat model is narrower than CVSS 8.5 suggests. Triggering this requires write access to the artifact store sufficient to craft artifact_path=’model$(cmd)’. Anyone with that level of access can also write a malicious python_model.py or modify the MLmodel file, both of which execute on model load. Landing the fix as defense-in-depth.”</p><h4>Assessment</h4><p>The “one of several pre-existing RCE paths” argument is technically sound. PR #23180 landed the fix as defense-in-depth. Copilot identified that shlex.quote (POSIX) should be replaced with mlflow.utils.string_utils.quote for Windows compatibility in both implementation and tests.</p><p><strong>PR #23180:</strong> <a href="https://github.com/mlflow/mlflow/pull/23180">https://github.com/mlflow/mlflow/pull/23180</a> <strong>Status:</strong> Merged in MLflow 3.12.0. Co-authored by @sreelim.</p><h3>Finding 5 — High: exec() on Deserialized Scorer Source</h3><p><strong>GHSA:</strong> GHSA-p68r-h9hx-w3hx | <strong>CWE:</strong> CWE-94</p><h4>Description</h4><p>MLflow deserializes scorer source code and passes it to exec(). A Databricks URI gate at mlflow/genai/scorers/base.py:524 is intended to restrict this path but is bypassable via environment variable injection, compromised config, or DNS hijacking of an existing databricks:// profile.</p><h4>Vendor Response</h4><p>“The bypass requires the attacker to control MLFLOW_TRACKING_URI on the victim’s machine. In all those cases the attacker already has environment or code control on the target — the exec() is then a redundant primitive, not a new attack surface. The threat model is circular.”</p><h4>Assessment</h4><p>The circular threat model argument is technically sound for the env var injection case. The DNS hijacking vector requires only network position rather than host compromise, which is a meaningfully different prerequisite; but I accept this is a design philosophy disagreement rather than a clear-cut vulnerability. No fix opened.</p><h3>Finding 6 — High: Hardcoded Default KEK Passphrase</h3><p><strong>GHSA:</strong> GHSA-xmgf-2fw4-whfc | <strong>CWE:</strong> CWE-798</p><h4>Description</h4><p>MLflow’s AI Gateway uses a hardcoded default KEK passphrase visible in the source repository to encrypt all gateway secrets, including API keys for connected LLM providers.</p><h4>Vendor Response</h4><p>“The default KEK passphrase is intentional behavior, documented at mlflow/utils/crypto.py:20–31. A runtime warning is emitted at crypto.py:169–174 when MLFLOW_CRYPTO_KEK_PASSPHRASE is unset. We may add a stricter check that refuses startup with the default passphrase when bound to a non-localhost interface.”</p><h4>Assessment</h4><p>I missed the existing runtime warning — that one’s on me and it meaningfully weakens the finding. The vendor’s proposed stricter startup check mirrors the remediation pattern applied to the admin password finding. Reasonable resolution. Tracked separately, no PR opened yet.</p><h3>Finding 7 — Critical: CORS Wildcard with Credentials Enabled</h3><p><strong>GHSA:</strong> GHSA-r7r2-rpww-m2cc | <strong>CWE:</strong> CWE-942</p><h4>Description</h4><p>MLflow’s init_security_middleware in mlflow/server/security.py registers CORS(..., origins=&quot;*&quot;, supports_credentials=True) when MLFLOW_SERVER_CORS_ALLOWED_ORIGINS=* is set. Flask-CORS works around the browser&#39;s wildcard-plus-credentials restriction by reflecting the request Origin header, enabling attacker-controlled origins to make authenticated cross-origin requests using the victim&#39;s session.</p><h4>Vendor Response</h4><p>“The configuration is admin opt-in via a clearly-named env var, default is closed, and the existing test documents the behavior as insecure. However, the wildcard-plus-credentials combination is a well-known web spec violation and we should refuse it at startup. We’ll land a fix that forces supports_credentials=False in the wildcard branch.”</p><h4>Assessment</h4><p>PR #23178 forces supports_credentials=False when wildcard origins are configured, logs a warning, and updates tests. After the initial commit, GitHub Copilot identified that the FastAPI middleware (mlflow/server/fastapi_security.py) retained the same allow_origins=[&quot;*&quot;] with allow_credentials=True combination in the ASGI path. A second commit addressed this gap, fixing both Flask and FastAPI paths and adding a regression test for the FastAPI case. Both commits are co-authored by @sreelim.</p><p>The scope of the fix expanded beyond what I originally reported because of Copilot’s automated review — the full fix is more complete as a result.</p><p><strong>PR #23178:</strong> <a href="https://github.com/mlflow/mlflow/pull/23178">https://github.com/mlflow/mlflow/pull/23178</a> <strong>Status:</strong> Merged in MLflow 3.12.0. Two commits — Flask path and FastAPI path. Co-authored by @sreelim.</p><h3>Finding 8 — Critical: Dockerfile Instruction Injection via Unsanitized base_image</h3><p><strong>GHSA:</strong> GHSA-m365-xv62-rcgx | <strong>CWE:</strong> CWE-94</p><h4>Description</h4><p>mlflow/projects/docker.py interpolates docker_env.image from an MLproject YAML file directly into a generated Dockerfile:</p><pre>dockerfile = f&quot;FROM {base_image}\nCOPY ...&quot;</pre><p>No validation is performed for newline characters. When build_image=True, an embedded newline introduces attacker-controlled Dockerfile instructions that execute during the Docker build with Docker daemon privileges (root by default on Linux).</p><p>Benign canary:</p><pre>docker_env:<br>  image: &quot;ubuntu:20.04\nRUN echo INJECTED &gt; /tmp/proof&quot;</pre><pre>mlflow run . --build-image</pre><p>Docker build logs confirm execution of the injected RUN instruction.</p><h4>Vendor Response</h4><p>“The premise is that an attacker authors a malicious MLproject and a victim runs mlflow run — build-image against it. In that threat model, the MLproject is already trusted input — the command: field is literally documented to execute arbitrary shell, so an attacker authoring the project has a much more direct vector than newline injection in base_image. Validating base_image doesn’t change the threat model.”</p><h4>Assessment</h4><p>This is the most substantive pushback in the set and worth engaging directly.</p><p>The command: field is explicit and documented as executable. Operators who review an MLproject file expect it to run commands. A newline in base_image is not explicit -- it is a hidden injection path that appears to be an image name. The attack surface is non-obvious in a way that command: is not.</p><p>In automated CI/CD pipelines, policy controls may restrict command: execution (OPA, image signing, artifact validation) without extending to image name validation. A newline-injected FROM instruction bypasses those controls silently.</p><p>That said, the vendor’s position is internally consistent and I accept the finding is closed. No fix opened.</p><h3>Finding 9 — High: Missing Authorization on CreateGatewaySecret</h3><p><strong>GHSA:</strong> GHSA-w2cj-3w5x-rm8w | <strong>CWE:</strong> CWE-862</p><h4>Description</h4><p>The MLflow AI Gateway CreateGatewaySecret endpoint does not perform authorization checks. Any authenticated user can create or overwrite gateway secrets.</p><h4>Vendor Response</h4><p>“Gateway secrets follow the same create-and-own pattern as CreateExperiment and CreateRegisteredModel: there is no before-request validator, and AFTER_REQUEST_PATH_HANDLERS grants the creator MANAGE on the new resource. The asymmetry with Get/Update/DeleteGatewaySecret is by design.”</p><h4>Assessment</h4><p>The create-and-own pattern is a coherent and consistently applied design choice across MLflow resources. I accept this response. No fix opened.</p><h3>What Happened in Twenty-Six Minutes</h3><p>After the first response at 8:15pm, the activity was remarkable:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JVKQhvvjH0tqPfJo68VsmQ.png" /></figure><p>Five pull requests, nine advisory reviews, substantive technical responses to each finding — in twenty-six minutes. The quality of the responses was high throughout. The vendor engaged with every finding on its technical merits, conceded where the code needed fixing even when declining vulnerability classification, and co-authored commits with the reporter on every PR.</p><p>GitHub Copilot’s automated review then identified additional gaps across the PRs:</p><ul><li>FastAPI middleware retained wildcard-plus-credentials CORS after PR #23178’s initial commit (addressed in second commit)</li><li>_str_optional() uses wrong quoting context for R parsing in PR #23179</li><li>shlex.quote (POSIX-only) used instead of mlflow.utils.string_utils.quote in PRs #23180 and #23181 (Windows compatibility gap)</li><li>_DEFAULT_ADMIN_PASSWORD constant in PR #23182 duplicates the value in basic_auth.ini and could drift over time</li></ul><p>These observations are legitimate and represent the kind of secondary review that makes fixes more robust before merge.</p><h3>On CVE Classification</h3><p>The vendor declined to assign CVEs to any of the nine findings. I disagree on two specifically.</p><p><strong>GHSA-7v72-hrvh-8fjh (hardcoded credentials):</strong> A finding that receives a bug fix PR co-authored with the reporter, marked for inclusion in the next release, is being treated as a vulnerability by the vendor’s own engineering process regardless of what the security team calls it. CVSS PR:N applies to the attacker’s privilege requirement, not the operator’s configuration prerequisite — these are different dimensions.</p><p><strong>GHSA-r7r2-rpww-m2cc (CORS):</strong> The fix expanded to two commits covering both Flask and FastAPI paths after Copilot identified the incomplete initial fix. A finding that requires fixing two separate middleware implementations is not a trivial configuration issue.</p><p>CVE IDs have been requested from MITRE and are pending assignment independent of vendor classification.</p><h3>Summary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*0Dx7mS0nPAZfQCLaDlXX2w.png" /></figure><p>CVE IDs pending MITRE assignment. This post will be updated when assigned.</p><h3>Disclosure Timeline</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/769/1*GAKhi12dzLd9MXNYB6BdLw.png" /></figure><h3>About the Author</h3><p><em>Sreejith Gopinath is a security researcher and currently works as an Engineering Manager at TRM Labs. He holds an MS in Computer Science from SDSU, an MS in Cybersecurity from NYU and has over a decade of experience in technology and finance. This research is part of an ongoing audit of open-source security tooling used in enterprise ML and data infrastructure.</em></p><p><em>GitHub: @sreelim</em></p><p><em>This is one in a series. Keep an eye out for the next one!</em></p><p><em>Full proof-of-concept payloads for findings 2, 3, 4, and 8 will be published 30 days after this post or upon vendor patch release, whichever comes first.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=caef89507bec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Found Six Security Issues in a CISA-Recommended DFIR Tool. Here’s What Rapid7 Said]]></title>
            <link>https://medium.com/@sreejith.gopinath/velociraptor-security-analysis-six-findings-and-a-conversation-with-rapid7-2b428a83069c?source=rss-97dd7846468b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b428a83069c</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[static-analysis]]></category>
            <dc:creator><![CDATA[Sreejith Gopinath]]></dc:creator>
            <pubDate>Thu, 07 May 2026 17:47:17 GMT</pubDate>
            <atom:updated>2026-05-17T22:03:19.747Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Velociraptor is a powerful open-source digital forensics and incident response (DFIR) platform used by security teams worldwide for endpoint visibility, artifact collection, and threat hunting. As a platform that frequently holds sensitive forensic evidence, credentials, and privileged access to endpoints, its own security posture warrants careful scrutiny.</p><p>This post documents six security findings identified through static analysis of the Velociraptor source tree at v0.76.4 and HEAD. For each finding I include Rapid7’s response, and where I disagree with their assessment I explain why. Where their response is well-reasoned I say so. The goal is to give operators a clear picture of the tradeoffs involved in deploying Velociraptor securely.</p><p>Rapid7 engaged professionally throughout this process and their documentation team has done substantive work on the hardening guide at https://docs.velociraptor.app/docs/deployment/security/. Operators deploying Velociraptor in production environments should read it carefully.</p><h3>Finding 1 — Critical: Secrets DEK Defaults to TLS Private Key</h3><p><strong>CWE:</strong> CWE-321 — Use of Hard-coded Cryptographic Key<br><strong>File:</strong> services/secrets/utils.go:44–66<br><strong>CVSS 3.1:</strong> AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:N — 8.1 High</p><h3>Analysis</h3><p>The Data Encryption Key (DEK) for the secrets store, which holds HTTP credentials, API tokens, SFTP passwords, and execve secrets is derived by default from ObfuscationNonce, falling back to Frontend.PrivateKey (the TLS private key) when Security.SecretsDek is not explicitly set:</p><p>go</p><pre>dek_str := config_obj.Security.SecretsDek<br>if dek_str == &quot;&quot; {<br>    dek_str = config_obj.ObfuscationNonce<br>}<br>if dek_str == &quot;&quot; {<br>    dek_str = config_obj.Frontend.PrivateKey<br>}</pre><p>Both the ciphertext (encrypted secrets in the filestore) and the key material (config file) are co-located. An attacker who exfiltrates the config file obtains both in a single operation. There is no key separation.</p><h3>Rapid7’s Response</h3><blockquote><em>“The docs suggest passing the DEK in the environment variable (which is the usual way to run containers for example). In fact it is possible to pass the entire config file itself as an environment variable to mitigate file compromise inside the container. In future we may suppose cloud based EKMS methods but this is not currently implemented so passing by ENV variable is a reasonable workaround. So this issue is actually documented and is working as intended.”</em></blockquote><h3>Assessment</h3><p>Rapid7’s response is technically accurate ; the env var path (env://) exists and is documented. The workaround is reasonable for containerized deployments.</p><p>However, <strong>a documented insecure default is not a secure default.</strong> The vast majority of Velociraptor deployments are probably not containerized and operators who have never read deployment/references/#security.secrets_dek will run with the TLS private key as their DEK indefinitely without knowing it. The secure-by-default principle means the safe path should require no configuration.</p><p><strong>Operator action required:</strong> Explicitly set Security.SecretsDek in your config, preferably via the env:// mechanism. Do not rely on the default.</p><p>yaml</p><pre>Security:<br>  SecretsDek: &quot;env://VELOCIRAPTOR_DEK&quot;</pre><p>Then set VELOCIRAPTOR_DEK to a randomly generated 32-byte key in your deployment environment.</p><h3>Finding 2 — High: AES-CBC Fixed IV in VQL Artifact Obfuscation</h3><p><strong>CWE:</strong> CWE-329 — Generation of Predictable IV with CBC Mode<br><strong>File:</strong> crypto/simple.go:41 (Encrypt), :89 (Decrypt)</p><h3>Analysis</h3><p>The Obfuscator used to hide VQL artifact names transmitted to clients uses self.key[:aes.BlockSize] as the AES-CBC IV — the same 16 bytes derived from ObfuscationNonce for every message. Encryption is fully deterministic: identical plaintext always produces identical ciphertext under the same config.</p><h3>Rapid7’s Response</h3><blockquote><em>“Therefore by definition, it has to be exactly the same for each message so it can be reversed by the server. i.e. there has to be a 1-&gt;1 mapping between obfuscated names and real names… I dont think this idea really works well — at least not yet, because there is a lot of information still present in the VQL queries which can not be obfuscated (like variable names). So this feature is not really anything to be relied on. We should improve the documentation around this.”</em></blockquote><h3>Assessment</h3><p><strong>Rapid7’s response is correct.</strong> The deterministic encryption is a design requirement for this use case, the server must be able to reverse the mapping. This is not a conventional encryption scheme and should not be evaluated as one.</p><p>The more important takeaway is their own concession: <strong>artifact name obfuscation should not be relied upon as a security control.</strong> Operators who believe artifact names are meaningfully hidden from endpoints should revise that assumption. Endpoint agents can potentially correlate ciphertexts across collections, and variable names within VQL queries remain visible regardless.</p><p><strong>Operator action required:</strong> Do not rely on artifact name obfuscation to protect sensitive operational details. Treat artifact names as potentially visible to endpoint agents.</p><h3>Finding 3 — High: Passwords Hashed with SHA-256, No Key Stretching</h3><p><strong>CWE:</strong> CWE-916 — Use of Password Hash with Insufficient Computational Effort<br><strong>File:</strong> services/users/set_user.go:150–164</p><h3>Analysis</h3><p>Built-in basic authenticator passwords are hashed with a single salted SHA-256 invocation. SHA-256 is not a password hashing function; it carries no work factor and runs at approximately 1–5 billion iterations per second on consumer GPUs. The salt prevents rainbow tables but not GPU-accelerated brute-force attacks. No rate limiting exists on the /api/v1 gRPC or HTTP basic-auth endpoints.</p><p>go</p><pre>func setPassword(user_record *api_proto.VelociraptorUser, password string) {<br>    salt := make([]byte, 32)<br>    _, err := rand.Read(salt)<br>    hash := sha256.Sum256(append(salt, []byte(password)...))<br>    user_record.PasswordSalt = salt[:]<br>    user_record.PasswordHash = hash[:]<br>}</pre><h3>Rapid7’s Response</h3><blockquote><em>“Velociraptor offers multiple authentication schemes for the web app. The simplest one is using Basic HTTP authentication. This is not the recommended method, actually for production systems we recommend using a SSO solution like oauth2 or OIDC or SAML… Since the Basic authentication must be verified on every HTTP interaction there is no key stretching to make it faster to hash the password. So this is the compromise — it is good enough for demo or evaluation type deployments but for production definitely use OIDC.”</em></blockquote><h3>Assessment</h3><p>Rapid7’s recommendation to use OIDC for production is correct and I agree with it. However, two issues remain.</p><p>First, <strong>“demo or evaluation” deployments running SHA-256 passwords routinely become production deployments.</strong> The absence of any in-product warning when basic auth is configured means operators have no signal to upgrade their auth posture. A startup warning or documentation callout at the basic auth configuration step would materially reduce risk.</p><p>Second, the performance argument (“no key stretching to make it faster”) is a false tradeoff. bcrypt with a moderate work factor adds approximately 100–300ms per authentication; negligible for a web UI that humans interact with, not a tight loop. The performance concern does not justify SHA-256 for password storage.</p><p><strong>Operator action required:</strong> Use OIDC, OAuth2, or SAML for any Velociraptor deployment that holds real forensic data or privileged access. Do not use basic authentication in production. If you must use basic auth, enforce strong passwords and restrict network access to the API endpoint.</p><h3>Finding 4 — High: InsecureNetworkTraceFile Disables All TLS Verification</h3><p><strong>CWE:</strong> CWE-295 — Improper Certificate Validation<br><strong>File:</strong> vql/networking/spy.go:61–98 (WebSocket), :140–172 (HTTP)</p><h3>Analysis</h3><p>When a client config has InsecureNetworkTraceFile set to any non-empty path, the framework replaces the normal TLS dialer with versions that set tls.Config{InsecureSkipVerify: true} with no VerifyConnection hook. All certificate validation for every outbound connection is silently disabled. There is no runtime enforcement preventing this setting from being active in production.</p><h3>Rapid7’s Response</h3><blockquote><em>“Yes this is a debugging tool that allows users to strip away the TLS encryption layer so they can see what is happening under the hood. It is not set by default and is documented… NOTE: Velociraptor has 2 layers of encryption — this setting bypasses the outer TLS layer to expose the underlying HTTP communications. Client-&gt;server communication still remain encrypted however.”</em></blockquote><h3>Assessment</h3><p>Rapid7’s clarification about the dual-layer encryption model is important and worth understanding. Velociraptor implements its own application-layer encryption on top of TLS. When InsecureNetworkTraceFile is set, the outer TLS layer is bypassed but the inner Velociraptor encryption layer remains active. This partially mitigates the MitM risk I described.</p><p>However, <strong>the outer TLS layer is what provides server authentication; </strong>verifying the client is talking to the legitimate Velociraptor server. Bypassing it enables a network-position attacker to impersonate the server for the TLS handshake, even if the inner encryption channel remains intact. The documentation note “Do not leave this configured in production!” is correct but has no runtime enforcement.</p><p><strong>Operator action required:</strong> Audit your client configs for InsecureNetworkTraceFile entries. This setting should never be present in production deployments. Consider adding a startup assertion that fails loudly if this is set in a non-debug build.</p><h3>Finding 5 — High: AllowedFsAccessorPrefix Reads Wrong Field</h3><p><strong>CWE:</strong> CWE-284 — Improper Access Control<br><strong>File:</strong> services/sanity/security.go:61–95</p><h3>Analysis</h3><p>The condition checks AllowedFsAccessorPrefix but the inner loop iterates over AllowedFileAccessorPrefix — a different field. When an operator sets AllowedFsAccessorPrefix, the allowlist tree is non-nil but empty, causing either a complete filestore access lockout or leaving all paths open depending on the intent.</p><p>go</p><pre>if len(config_obj.Security.AllowedFsAccessorPrefix) &gt; 0 {<br>    allowed_tree = utils.NewPrefixTree(false)<br>    for _, allowed := range config_obj.Security.AllowedFileAccessorPrefix {<br>        // iterates wrong field — tree remains empty<br>    }<br>}</pre><h3>Rapid7’s Response</h3><blockquote><em>“I don’t believe that the AllowedFsAccessorPrefix bug is actually a CVE because most people would not use this field. The documentation actually suggests using the denied_file_accessor_prefix field… This is a relatively recent security control which is actually an additional control to the usual user ACLs — only administrators would normally be able to read from the filestore anyway.”</em></blockquote><h3>Assessment</h3><p>Rapid7 confirmed this is a bug. Their CVE denial is based on low exploitability; the field is rarely used and admins already have filestore access via ACLs. This is a reasonable position.</p><p>That said, <strong>a security configuration option that silently does nothing is worth documenting explicitly.</strong> Operators who set AllowedFsAccessorPrefix believing it protects them are unprotected without knowing it. The one-character field name fix is trivial; I would encourage the team to patch it regardless of CVE status, and to add a log warning when AllowedFsAccessorPrefix is set without a corresponding AllowedFileAccessorPrefix entry.</p><p><strong>Operator action required:</strong> Use denied_file_accessor_prefix (not AllowedFsAccessorPrefix) to restrict filestore access. The default denied prefixes — acl, backups, config, orgs, secrets, users — are applied correctly and protect the most sensitive paths.</p><h3>Finding 6 — High: SSRF via Unix Domain Socket Support in http_client</h3><p><strong>CWE:</strong> CWE-918 — Server-Side Request Forgery<br> <strong>File:</strong> vql/networking/http_client.go:628–649</p><h3>Analysis</h3><p>The http_client VQL plugin accepts a URL syntax of /&lt;socket_path&gt;:unix/&lt;http_path&gt; that routes HTTP requests over local Unix domain sockets, gated only by the NETWORK ACL permission. On Docker hosts or Kubernetes nodes, a user with NETWORK permission can reach /var/run/docker.sock:</p><p>vql</p><pre>SELECT * FROM http_client(url=&quot;/var/run/docker.sock:unix/v1.41/containers/json&quot;)</pre><h3>Rapid7’s Response</h3><blockquote><em>“So the feature is there by design. The issue is only if it is safe to allow this on the server (e.g. using it in a notebook or server artifact)… using the plugin requires the network permission which is normally only given to admin — so if you can use the usual ACLs to control access to the plugin.”</em></blockquote><blockquote><em>Additional mitigations offered:</em></blockquote><ul><li>Disable the plugin completely via server config</li><li>Use security.vql_must_use_secrets to restrict http_client to allowlisted URLs</li><li>Note that client agents have no ACL restrictions at all</li></ul><h3>Assessment</h3><p>Rapid7’s response is technically sound. The Unix socket feature is intentional and documented; they use it themselves in Docker artifacts. The risk is real but gated behind NETWORK permission which defaults to admin-only.</p><p>The key operator takeaway is the <strong>client-side caveat Rapid7 raised</strong>: Velociraptor clients have no ACL restrictions. A VQL artifact deployed to an endpoint agent can call http_client with a Unix socket URL without any permission check. This is a more significant risk than the server-side scenario and worth explicit documentation.</p><p><strong>Operator action required:</strong></p><ul><li>On shared servers, consider disabling http_client entirely or using security.vql_must_use_secrets to allowlist permitted URLs</li><li>Review server artifacts and notebooks that use http_client for Unix socket usage</li><li>On Docker/Kubernetes deployments, ensure the Velociraptor server process does not have access to /var/run/docker.sock</li><li>Be aware that client-deployed artifacts have no ACL restrictions on http_client usage</li></ul><h3>Deployment Hardening Summary</h3><p>Based on this analysis and Rapid7’s responses, the following hardening steps are recommended for production Velociraptor deployments:</p><p><strong>Authentication</strong></p><ul><li>Use OIDC, OAuth2, or SAML — never basic auth in production</li><li>If basic auth is unavoidable, restrict API endpoint network access</li></ul><p><strong>Secrets and Encryption</strong></p><ul><li>Explicitly set Security.SecretsDek via the env:// mechanism</li><li>Generate a unique 32-byte random DEK per deployment</li><li>Do not rely on artifact name obfuscation as a security control</li></ul><p><strong>Network and TLS</strong></p><ul><li>Audit all client configs for InsecureNetworkTraceFile — must be absent in production</li><li>On Docker/Kubernetes, isolate the Velociraptor server from Docker socket access</li></ul><p><strong>Access Control</strong></p><ul><li>Use denied_file_accessor_prefix (not AllowedFsAccessorPrefix) for filestore restrictions</li><li>Default denied paths (acl, backups, config, orgs, secrets, users) are applied correctly</li><li>On shared servers, consider disabling http_client or using vql_must_use_secrets</li><li>Review server artifacts and notebooks for Unix socket usage in http_client</li></ul><p><strong>Full hardening documentation:</strong> <a href="https://docs.velociraptor.app/docs/deployment/security/">https://docs.velociraptor.app/docs/deployment/security/</a></p><h3>Disclosure Timeline</h3><ul><li>May 2, 2026: Findings reported to Rapid7 via coordinated disclosure</li><li>May 6, 2026: Rapid7 responded addressing all six findings</li><li>May 7, 2026: Follow-up exchange on finding 5 CVE and blog post notification</li><li>May 7, 2026: This post published with Rapid7’s responses included</li></ul><h3>About the Author</h3><p>Sreejith Gopinath is an Engineering Manager at TRM Labs, with former stints at Palo Alto Networks and Apple. He is an open-source security researcher in his own time. Sree holds an MS in Computer Science from SDSU and an MS in Cybersecurity from NYU with several years of professional experience in cybersecurity. This research is part of an ongoing audit of open-source security tooling used in enterprise environments.</p><p>GitHub: <a href="https://github.com/sreelim">@sreelim</a></p><p><em>This post is one in a series of security audits of open-source security platforms. Additional findings across other platforms will be published separately.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b428a83069c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cybersecurity Battlefield Is Now Model vs. Model]]></title>
            <link>https://medium.com/@sreejith.gopinath/the-cybersecurity-battlefield-is-now-model-vs-model-1317880580b3?source=rss-97dd7846468b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1317880580b3</guid>
            <category><![CDATA[mytho]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sreejith Gopinath]]></dc:creator>
            <pubDate>Sun, 12 Apr 2026 22:27:44 GMT</pubDate>
            <atom:updated>2026-04-12T22:30:05.046Z</atom:updated>
            <content:encoded><![CDATA[<p>You have already read the sandwich story. We will not revisit it.</p><p>What matters is not the anecdote. What matters is what comes next, and most of what has been written about Mythos this week stops exactly where the interesting questions begin.</p><p>Forty organizations have Mythos Preview access right now. They are among the most sophisticated security operations on the planet: Amazon, Apple, Microsoft, CrowdStrike, Google, Palo Alto Networks, JPMorgan Chase, the Linux Foundation. For a limited and shrinking window of time, they hold a capability edge that is unlike anything the security industry has seen before.</p><p>That edge will not hold. It cannot. AI cybersecurity capability will democratize, the same way every significant security capability before it eventually democratized. The question worth asking is not what Mythos did during testing. The question is what the world looks like when the forty becomes four thousand, and then four million.</p><p><strong>The Backward-Looking Conversation</strong></p><p>The security industry’s response to Mythos has been almost entirely backward-looking. Decades-old vulnerabilities found in OpenBSD. A sixteen-year-old bug in FFmpeg that automated tools had touched five million times without catching. The compliance frameworks we rely on, SOC 2, FedRAMP, NYDFS Part 500, calibrated against human-speed threat actors and audit cycles measured in quarters, now exposed as structurally inadequate against a model that develops working exploits on first attempt in 83% of cases.</p><p>All of that is true and worth understanding. But it addresses only the accumulated liability of past code. The more uncomfortable question is forward-looking.</p><p>What about the bugs yet to be written?</p><p>We are all today writing code with AI assistants that may themselves introduce novel vulnerability patterns. The attack surface is not a fixed artifact of technical debt. Like the universe itself, it is expanding right now, today, at a pace that has never been higher, by teams that are moving faster than ever before. Mythos-class models will find those vulnerabilities trivially. We have not begun to reckon with what that means for software being written in 2026.</p><p>The two curves, rate of vulnerability introduction and rate of exploitation capability, are moving together. That is a different and more serious problem than either curve moving alone.</p><p><strong>Three Classes of Defenders</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dm7ZwZB-ilpJZ-kZZwEmyg.jpeg" /><figcaption>The three defender classes</figcaption></figure><p>The Mythos moment has created a new taxonomy that every security organization needs to understand and locate itself within.</p><p>The first class is defenders who have Mythos. Forty organizations. Unprecedented capability. A window to harden attack surfaces against vulnerabilities that human reviewers and automated tools missed for decades. This class is set apart not just by access to a tool but by a fundamentally different threat model, one calibrated to what AI-speed offensive capability actually looks like.</p><p>The second class is defenders who do not have Mythos but know they need equivalent capability. They are watching the forty, absorbing the published research from Project Glasswing, rebuilding their threat models, and making decisions about posture before the urgency becomes undeniable. They will not have the same head start but they are in motion.</p><p>The third class is the largest and the most exposed. Defenders who do not yet realize the baseline has moved. Organizations still treating a clean quarterly scan as meaningful assurance. Security programs built for the threat environment of three years ago, calibrated against human-speed adversaries, confident in compliance postures that were never designed for what is coming.</p><p>When Mythos-class capability reaches threat actors, and Anthropic’s own estimate is six to eighteen months, the third class becomes the primary target. Not because they are the most valuable targets. Because they are the most accessible ones.</p><p>The work of the next twelve months is moving organizations from the third class to the second. The window to do that deliberately, before the pressure becomes acute, exists right now and is closing.</p><p><strong>The Inevitable Democratization</strong></p><p>Every significant security capability in history has followed the same arc. It begins as the exclusive domain of the most sophisticated actors: nation-states, elite research teams, well-funded enterprises. It becomes a commercial product. It becomes a commodity. It ends up in the hands of everyone, attackers included.</p><p>Mythos will follow this arc. The question is not whether AI-level vulnerability discovery and exploit development will be widely available. It is when, and whether defenders or attackers get there first in meaningful numbers.</p><p>The forty organizations with access today are not a permanent elite. They are the early access cohort of a capability that will, within years, be table stakes for any serious security operation. The organizations building the posture, the processes, and the institutional knowledge to operate with AI-augmented security today are the ones that will not be caught flat-footed when that capability becomes standard.</p><p>Democratization of defensive capability is necessary and inevitable. The concern is that offensive capability democratizes on the same timeline, or faster. History suggests it usually does.</p><p><strong>The Battle of the Models</strong></p><p>The trajectory of all of this points toward something the industry has not fully confronted yet.</p><p>If Mythos-class models become standard defensive infrastructure, and they will, the next generation of offensive capability will not simply look for software vulnerabilities. It will probe the blind spots and reasoning patterns of the defensive models themselves. The target shifts from the infrastructure to the intelligence defending it.</p><p>Adversarial models trained specifically to operate beneath the detection threshold of Mythos-class defensive systems. Offensive models that fingerprint the defensive model’s behavior and craft attacks designed to evade its specific reasoning patterns. Models attempting to subvert the intelligence of other models rather than the systems those models protect.</p><p>This is not fantastical thinking. It is the logical extension of a pattern that has repeated throughout security history. Every defensive technology eventually becomes the target. Antivirus signatures get evaded. Intrusion detection rules get fingerprinted and avoided. The same dynamic will play out at the model layer, at a speed and scale that human defenders cannot match directly.</p><p>What follows is a battlefield that looks less like human analysts using AI tools and more like AI systems contesting each other at machine speed, with humans setting objectives, reviewing outcomes, and making strategic decisions but no longer executing in the loop.</p><p>This raises a question the industry needs to start answering now, before the battlefield is established. If the contest becomes model versus model, what is the role of the human defender? Strategist. Arbitrator. Ethical governor. These are meaningful roles. But they are different roles, and the organizations that thrive in that environment will be the ones that started building toward it before it arrived.</p><p><strong>What This Requires</strong></p><p>Not panic. Not a new vendor. A different posture built on honest assumptions about where the threat environment is going.</p><p>Continuous visibility over periodic assessment. The attack surface is not stable between scan cycles. It is being written right now.</p><p>Prioritization calibrated to AI-speed exploitation. The medium-severity finding on an internet-facing asset with public proof-of-concept code is not a thirty-day remediation item anymore. It is an immediate one.</p><p>Security programs designed to operate in a world where the defensive tool and the offensive tool are increasingly the same underlying capability, pointed in different directions by different actors with different intentions.</p><p>And perhaps most importantly: start thinking now about how your organization operates when the contest is model versus model. The organizations that will navigate that environment well are not necessarily the ones with the largest budgets. They are the ones that saw it coming and built accordingly.</p><p>The forty have a head start. The window to close that gap is open. It will not stay open.</p><p><strong><em>About the Author</em></strong></p><p><em>At TRM Labs, I lead teams building blockchain intelligence and AI-powered threat detection capabilities used by federal law enforcement to disrupt financial crime, ransomware networks, and darknet operations. That work sits at the intersection of AI capability and adversarial behavior, exactly the intersection that Mythos makes impossible to ignore.</em></p><p><em>Prior to TRM, I was part of the team that built Cortex Xpanse at Expanse, acquired by Palo Alto Networks. Cortex Xpanse is the attack surface management platform now used by the Department of Defense, CISA, and other federal agencies to maintain continuous visibility into internet-facing assets. The core insight behind that work remains true: you cannot defend what you cannot see, and most organizations cannot see nearly enough of their own attack surface.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1317880580b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>