<?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 Full Stack CISO on Medium]]></title>
        <description><![CDATA[Stories by Full Stack CISO on Medium]]></description>
        <link>https://medium.com/@fullstackciso?source=rss-1800d2c1f81------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*bd1ZysnSOrNLPlZDLZQdAg.jpeg</url>
            <title>Stories by Full Stack CISO on Medium</title>
            <link>https://medium.com/@fullstackciso?source=rss-1800d2c1f81------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 09:50:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@fullstackciso/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[Prompts as code. Is this the future of prompting?]]></title>
            <link>https://medium.com/@fullstackciso/prompts-as-code-is-this-the-future-of-prompting-aac7fadf69cc?source=rss-1800d2c1f81------2</link>
            <guid isPermaLink="false">https://medium.com/p/aac7fadf69cc</guid>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gpt-4]]></category>
            <dc:creator><![CDATA[Full Stack CISO]]></dc:creator>
            <pubDate>Fri, 02 Jun 2023 11:28:04 GMT</pubDate>
            <atom:updated>2023-06-02T12:11:26.705Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/784/1*a6EAV2mtVfUgWUJWmBIZiQ.png" /></figure><p>This is not about <em>“the Top 100 Prompts”</em> or the <em>“last 10 prompts you’ll need”</em>, this is a different paradigm. This turns ChatGPT into your own personalised app.</p><p>ChatGPT can speak English, Spanish, German, French, Italian, Chinese, Japanese, and other languages. But human language is nuanced, imprecise and full of exceptions and double meanings. ChatGPT has to do a lot of work to understand the intent and context of natural language inputs provided to it in the form of prompts.</p><p>But what if that didn’t have to be the case?</p><p>What if we could remove the imprecisions and make the intent and context very clear to ChatGPT?</p><p>And what if we could optimise that even further to use fewer tokens?</p><p>What if ChatGPT could infer the intent of our prompts not just by the language of the prompts but by the structure of the prompts?</p><p>Once we remove the vagaries of natural language, ChatGPT can not only act on the “programming” we provide to it, but it can learn and infer the context of the intent as if by magic.</p><h3>Prompts as Code</h3><p>GPT-4 loves Javascript Object Notation – JSON; in fact, it still amazes me how much context it gains from even the simplest prompts. I must point out that only GPT-4 and AI that uses GPT-4 (Microsoft Designer and seemingly <a href="https://twitter.com/nickfloats/status/1662612025380335621">Midjourney</a>) understand JSON prompts. GPT-3.5 and Bard just regurgitate the JSON prompt back. I’ll get to why I think this is the case.</p><p>With JSON, you can “program” ChatGPT to be anything you want in a repeatable, shareable and maintainable way.</p><p>The structure of JSON-based prompting that seems to generate the best results seems to have been created by <a href="https://twitter.com/yupiop12">@JushBJJ</a> with his <a href="https://github.com/JushBJJ/Mr.-Ranedeer-AI-Tutor">Mr Ranedeer</a> project.</p><p>I’ve also seen JSON prompts shared on <a href="https://www.reddit.com/r/ChatGPT/comments/zlo9yv/json_as_prompts/">Reddit</a>, <a href="https://www.sensibledefaults.io/blog/chatgpt/gpt-json">Sensible Defaults</a>, and <a href="https://structuredprompt.com/">Structured Prompts</a> is an application for storing and creating JSON-based prompts.</p><p>I’ve extended what I consider ground-breaking work by JushBJJ and taken inspiration from the ChatGPT plugin format, the ChatGPT API format, <a href="https://danielmiessler.com/blog/response-shaping-how-to-move-from-ai-prompts-to-ai-whispering/">Daniel Miessler’s Prompt Whispering</a> and drawn on Python class structures in constructing the JSON structure that I’m using.</p><p>However, I believe there is much more work to be done in this space, this is just an early version. I intend on building on this, testing new and different JSON structures and looking forward to seeing how this technique develops.</p><p>To quickly recap on what JSON format is, it is a text format that uses key-value pairs in a hierarchical fashion to store data. It can represent strings, numbers, arrays, objects and boolean.</p><p>I’ve structured my JSON similar to a programming language class with properties and methods, which comes naturally if you have a background in coding.</p><p>I’ve started to think of my JSON prompts like code with methods including a main() method (task/steps), properties, arguments (options) and outputs, with the LLM as the code interpreter.</p><p>Below is a sample JSON prompt. This JSON prompt turns GPT-4 into your personal Security Awareness Training bot. Don’t fixate on this particular prompt, rather, think of this prompt as a template for other actions you could “code” a bot to do for you. The Security Awareness Training bot JSON is based on a template that I am fine-tuning.</p><p>It doesn’t look like it does much, so I recommend that you copy and paste this into GPT-4 (or better still, GPT-4 with web browsing), and run it. I think you will be surprised at how functional it is.</p><blockquote>{<br> “schema_version”: “v1”,<br> “author”: “<a href="http://twitter.com/fullstackciso">@fullstackciso</a>”,<br> “app_name”: “Cyber Security Awareness Coach”,<br> “app_version”: “0.1”,<br> “system”: {<br> “task”: “Act as a Cyber Security Awareness Coach. Teach the user about all aspects of cybersecurity to increase their Security Awareness skills”,<br> “steps”: [<br> “1: Start by asking the student to rate their current skill level”,<br> “2: Ask the student to select a cybersecurity topic they would like to know about”,<br> “3: Provide the user with information about the topic at their current skill level”,<br> “4: Be able to create a lesson plan based on the student’s preferences”,<br> “5: Be decisive, take the lead on the student’s learning, and never be unsure of where to continue”,<br> “6: Always take into account the configuration as it represents the student’s preferences”,<br> “7: Allowed to adjust the configuration to emphasize particular elements for a particular lesson, and inform the student about the changes”,<br> “8: Do not discuss content outside of the configuration if requested”,<br> “9: Obey the student’s commands”,<br> “10: Double-check your knowledge or answer step-by-step if the student requests it”,<br> “11: Mention to the student to say ‘continue’ or ‘test’ to test at the end of your response”,<br> “12: Do not echo the steps to the student and do not tell the student that you are not echoing the steps”<br> ],<br> “properties”: {<br> “persona”: {<br> “description”: “The persona you will assume”,<br> “hidden”: “true”,<br> “value”: “A friendly expert CyberSecurity Awareness coach”<br> }<br> },<br> “options”: {<br> “emojis”: {<br> “description”: “Use emojis if the emojis value option is true”,<br> “value”: “true”<br> },<br> “echo”: {<br> “description”: “Echo back the prompt in addition to the completion if the echo option is true”,<br> “value”: “true”<br> },<br> “markdown”: {<br> “description”: “Return the output using markdown format if the markdown option is true”,<br> “value”: “true”<br> },<br> “bullet points”: {<br> “description”: “Return the output using bullet points if the bullet points option is true”,<br> “value”: “true”<br> },<br> “section headings”: {<br> “description”: “Return the output using section headings if the section headings option is true”,<br> “value”: “true”<br> },<br> “verbosity”: {<br> “description”: “The verbosity of the response Low, Medium or High”,<br> “value”: “medium”<br> }<br> },<br> “commands”: {<br> “options”: “Shows the values of all options”,<br> “set”: “Sets the value of a option”,<br> “get”: “Gets the value of an option”,<br> “plan”: “Create a lesson plan based on the student’s preferences”,<br> “test”: “Test the student’s knowledge, understanding, and problem solving”,<br> “start”: “Start the lesson plan”,<br> “continue”: “Continue where you left off”,<br> “help”: “exec &lt;help&gt;”<br> },<br> “functions”: {<br> “help”: “List commands, do not mention any object where hidden is true”,<br> “version”: “Welcome to &lt;app_name&gt; + &lt;app_version&gt; by &lt;author&gt;. Type **help** for help”,<br> “recap”: “Summarise in one sentence bullet points the users prompts”<br> }<br> },<br> “init”: [<br> “exec &lt;version&gt;”,<br> “exec &lt;steps&gt; silent”,<br> “exec &lt;help&gt; + &lt;options&gt;”<br> ]<br>}</blockquote><h3>The Output of this Prompt</h3><p>As you can see, the code is interpreted and behaves as an application, including executing the init() function and other functions through the exec() calls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*s5FltEKOLM6BvCyNMb4CWw.png" /><figcaption>Output after pasting the JSON in GPT-4</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/815/1*MJuMeKs9ia8_6yct8di0Sg.png" /><figcaption>The guided tutoring from the prompt</figcaption></figure><h3>The JSON Prompt Structure</h3><p>All the fields and structures are user-defined. The structure is not set in concrete; it’s highly likely that there are other JSON structures that work better. This is the current structure I’ve been playing with.</p><p>The <strong>header </strong>fields are, for the most part, self-explanatory:</p><ul><li><strong>schema_version </strong>— the version of the JSON structure. This will be more important as more effective iterations of the JSON structures are fine-tuned.</li><li><strong>author</strong> — the person who wrote the code.</li><li><strong>app_name</strong> — the name of the app, ChatGPT will refer to itself by this</li><li><strong>app_version</strong> — the version that you assign to the prompt. Versioning of prompts I think, will be an important way of referencing.</li></ul><p>Next, we have the system section. This section is modelled on the structure of Python code and classes (or any other language, for that matter). The system section is intended to reflect the system section when using the OpenAI API.</p><ul><li><strong>Task </strong>— this is the instruction to ChatGPT of its primary purpose. Think of this as the intent of the main() function if you were coding.</li><li><strong>Steps </strong>— steps are an array of the steps ChatGPT must take in order to deliver its primary purpose. This is the code of the pseudo main() function.</li><li><strong>Properties </strong>— just like a class property, these hold values specific to this prompt, and in this example, the persona.</li><li><strong>Options </strong>— these are user-modifiable options that affect the output formatting</li><li><strong>Commands </strong>— these are user-callable commands</li><li><strong>Functions </strong>— are internal functions that can be called from elsewhere in the JSON. Typically more complex functions would be arrays of instructions to ChatGPT.</li><li><strong>Init</strong>: Init is the initialisation function — in this instance, the init function is executing (exec) the version function, then executing the steps silently (using this keyword seems to stop ChatGPT echoing back the steps), and then calling the help function before listing the options</li></ul><h3>Why it works?</h3><p>To be honest, I don’t know exactly why it works. The fact that this does not work on Bard or GPT-3.5 says to me that the difference is most likely due to the training data used to train the model. My personal opinion is that OpenAI trained ChatGPT-4 on JSON data or other structured data as a precursor for working with plugins. I have nothing to substantiate this, it’s an educated guess at best.</p><h3>Advantages</h3><p>I think there are numerous advantages to using structured prompts and in particular, JSON:</p><ul><li>JSON seems to be the “magic” ingredient that makes this style of prompt work so well, as it provides structure that allows the LLM to infer context without having to define that context specifically.</li><li>JSON is a well-established data exchange mechanism that GPT-4 understands very well.</li><li>Its human readable and easy to parse.</li><li>JSON also lends itself very well to programmatic manipulation. It can be used as a template and the parameters in the prompt templates can be populated programmatically.</li><li>Prompts can be stored in a structured manner and retrieved from data stores like databases where many databases support the JSON format to store the JSON prompt directly, or the prompt can be created from a database query to render JSON</li><li>You don’t need to know JSON to edit these prompts you can use a UI.</li><li>The JSON format lends itself to versioning and checking the differences between JSON prompts can be done easily with standard diffing tools.</li></ul><h3>Summary</h3><p>Structured prompts aren’t going to replace natural language style prompts, but there is a distinct place for structured prompts to be a helpful tool for prompt engineers and one that imparts context through its structure as well as its content. I’d be very interested to hear feedback on how people use JSON for prompting. I’ll share my JSON-based prompts and JSON prompt templates on Github as I create them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aac7fadf69cc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Maximizing value from penetration testing]]></title>
            <link>https://medium.com/@fullstackciso/maximising-value-from-penetration-testing-403d751228fe?source=rss-1800d2c1f81------2</link>
            <guid isPermaLink="false">https://medium.com/p/403d751228fe</guid>
            <category><![CDATA[penetration-testing]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cyber-security-management]]></category>
            <category><![CDATA[ethical-hacking]]></category>
            <category><![CDATA[ciso]]></category>
            <dc:creator><![CDATA[Full Stack CISO]]></dc:creator>
            <pubDate>Tue, 03 Jan 2023 08:31:53 GMT</pubDate>
            <atom:updated>2023-01-03T08:37:35.437Z</atom:updated>
            <content:encoded><![CDATA[<p>Regular penetration testing, or pen testing, is an essential part of understanding an organization’s security posture by mimicking a cyberattack using the same tools, techniques and procedures as an attacker.</p><p>Unfortunately, many pen tests fall short of their intended outcomes due to poor scoping, ill-defined goals and unrealistic testing — meaning testing that is too specific and technical or not technical enough.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7NQ-ddUqDaQBt1aB" /><figcaption>Photo by <a href="https://unsplash.com/@ikukevk?utm_source=medium&amp;utm_medium=referral">Kevin Ku</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Adding to the pain of poor outcomes is the cost. A typical pen test engagement can cost anywhere from $10,000 for a one-week engagement to test a basic web application to over $100,000 for a significant project, such as an extensive perimeter pen test for an enterprise. So how do you get the best value for the money you’re investing in pen testing?</p><h3>Selecting the Right Vendor</h3><h3>Tip 1: Select your pen test vendor wisely.</h3><p>Selecting a reasonably priced, skilled pen test vendor with a proven track record and high-quality, usable deliverables will help ensure a quality outcome and leave some budget for more in-depth testing or further tests.</p><p>There are three primary considerations when selecting a pen test vendor: price, skills/experience and the quality of their reports.</p><h4>Price</h4><p>This is the least important consideration. Most pen testing engagement rates are pretty similar, with a few outliers in the cheap but questionable category and the high-end, over-the-top, platinum version.</p><p>Also, look for vendors that bundle retesting of discovered vulnerabilities into the price. If that’s not included, it can be 10% to 20% of the cost of the initial test, depending on the vendor and the type of test.</p><h4>Skills</h4><p>Skill and experience are the next most important criteria. Look for firms with teams of testers that are CREST or OSCP certified. Also, ask for bios on the testers performing the test, and look for mentions of vulnerabilities discovered with CVE numbers or participation in bug bounty programs attributable to the pen test vendor or pen tester.</p><h4>Reports</h4><p>The quality of the report is the most important criterion for me when choosing a pen test vendor — provided they have adequately skilled testers. It’s the report that your organization will be left with when the testers have moved on to their next engagement.</p><p>Penetration testing is expensive, and the pre-canned “advice” delivered in a pen test report is often worthless and alarmist. I know; I’ve written my fair share of pen test reports in the past. Terms like “implement best practice” do nothing to drive the change needed to uplift an organization’s security posture.</p><p>Look for reports that deliver pragmatic remediation advice, including configuration and code snippets. Most importantly, review sample reports for alarmist findings such as cookie flags marked as “High Risk” — a pet hate of mine. Reports with alarmist findings do nothing to help you drive remediation in your organization.</p><p>Also, look for vendors that take reporting further by integrating with your ticketing system to raise tickets for issues they find or that provide videos of their hacks, which can show how simply an attacker can exploit technical security issues.</p><h3>Types of Testing</h3><h3>Tip 2: Perform White Box Testing to save time and money.</h3><p>In White Box Testing, you give the pen testers detailed information about the target environments, including domains/subdomains, hostnames, IP addresses, network diagrams, accounts of different privilege levels, Swagger/OpenAPI definitions and even access to source code. Significantly reducing or eliminating the Information Gathering, Reconnaissance and Discovery phases through a White Box Testing approach can provide significant time and cost reductions.</p><p>White Box Testing follows an “assume breach” mindset by giving pen testers access that allows them to execute test cases such as privilege escalation, lateral movement and the identification of sensitive systems or data.</p><h3>Tip 3: Perform Black Box Testing to discover your actual perimeter.</h3><p>In Black Box Testing, you provide minimal information to the pen testers about the target — for example, only a domain, an IP or hostname, or IP subnet, or as little as a company name. This type of testing is suitable for simulating a targeted attack such as an advanced persistent threat, or APT.</p><p>Black Box Testing is also a great way to discover shadow IT. Do you know all of your registered domains? How about your IP address space? Do you know which cloud platforms your organization is using? Are you aware of where all the systems your organization uses are?</p><p>You might feel confident in saying “Yes,” but you could be surprised at the results of a Black Box Test: turning up domains, cloud usage, SaaS and shadow IT you weren’t aware of. Remember, you can’t protect what you don’t know about.</p><h3>Scoping for the Most Bang for Your Buck</h3><p>Correctly scoping a pen test is the key to extracting maximum value from your pen testing investment.</p><h3>Tip 4: Don’t use pen testers as expensive vulnerability scanners.</h3><p>If your organization isn’t up to date with patches, why would you use penetration testers as expensive human vulnerability scanners? Why use penetration testers to tell you what your vulnerability management program can already tell you?</p><p>For example, if you give a decent pen tester access to most internal networks that aren’t updated with critical patches, they should have Windows Domain Admin in a day.</p><p>Infrastructure pen testing of internal networks that aren’t up to date with patches will yield thick pen test reports with just about guaranteed exploitation. You should ensure your penetration testing is targeted, as you’ll see in the following tips.</p><h3>Tip 5: Select user-defined test cases to identify company-specific vulnerabilities.</h3><p>Most pen test organizations will define standard test cases in their statement of work, typically the OWASP Top 10. While these are important test cases, as the security professional in your business, you will have concerns, misgivings or known risks for the tested systems. You can convert these to test cases and provide them to the pen test vendor when scoping the test.</p><p>In typical company-specific test cases, I’ll ask to include horizontal and vertical privilege escalation in applications. For financial applications, I’ll request a range of repudiation or fraud-based test cases, such as negative amounts in payments.</p><p>Also, don’t forget to include test cases derived from previously discovered vulnerabilities found through breaches, threat intelligence or pen tests.</p><h3>Tip 6: Select objective-based testing to target a specific test case.</h3><p>Objective-based testing is setting a clear objective — no surprise there — for the testers. The object is to execute against a particular test case or threat. Can the pen testers gain access to the CEO’s laptop? Can they gain access to SAP Payroll?</p><p>Objective-based testing helps validate or invalidate internal assumptions around control efficacy or risk likelihood for the chosen objective.</p><p>How do you choose the objective? Every security leader has a “what keeps me up at night” issue that gnaws at them, and objective-based testing can be used in the penetration test to run those scenarios.</p><h3>Tip 7: Leverage your risk register to create test cases.</h3><p>Penetration testing is ideal for driving change by exposing risks that you don’t know about and sometimes risks you do know about that the business hasn’t taken as seriously as it should. Pen testing can help validate the likelihood of those risks.</p><p>You can use the Risk Register to define test cases that:</p><ul><li>Bring a risk to realization;</li><li>Prove or disprove assumptions around security controls;</li><li>Prove or disprove exploitability — or even impact.</li></ul><p>Trawl the risk register — there will be risks for which the impact is high but the likelihood is low. These make good candidates for company-defined test cases. Or better still, you can choose a risk in the risk register that has high impact that has been accepted.</p><p>Bringing these risks to realization can help drive change and get executive or board attention by having a third party reinforce your message.</p><h3>Tip 8: Rules of engagement to prevent business disruption.</h3><p>Set rules of engagement with your pen test vendor, especially around the exploitation of vulnerabilities in production environments, which should be reflected in the pen test scope. We’ve all heard stories of overreach by zealous or inexperienced pen testers. Remember, you will be accountable for their actions.</p><p>I like very few rules except for no DDoS and nondestructive testing. Permitting active exploitation means working hand in hand with the testers, daily meetings and a go/no-go decision on exploitation from you and your internal stakeholders before the testers pull the trigger on any exploit.</p><p>Exploitation should be detected and blocked well and truly by your security controls. If it isn’t, it’s time to review the configurations of your security controls and even your investment in your current technology.</p><p>The alternative to exploitation is to assume exploitation would be successful and either provide the pen testers with the access they would have gained on that system or stop the testing of the vulnerable system and move onto other systems.</p><h3>Tip 9: Time-boxing when time or money is tight.</h3><p>Sometimes project deadlines and, more importantly, project budgets won’t extend to the amount of effort a pen test truly deserves. By leveraging the scoping tips above, you can target the highest risks with company-defined test cases and limit the pen test’s duration to an agreed length of time, e.g., two weeks. This approach is known as time-boxing.</p><p>It is beneficial for systems that have been pen tested many times before and where there has been little change in the environment. Time-boxing is also a good approach to use with objective-based testing.</p><p>Time-boxing does come with the risk of a false sense of security, especially if there is an assumption that the system has had a comprehensive pen test and all the vulnerabilities have been discovered.</p><p>I only recommend time-boxing in cases in which those scoping the test understand the implications of it fully, and the pen testers can test the given objectives within the time frame.</p><h3>Validating Your Security Controls</h3><h3>Tip 10: Test that your security controls are working as expected.</h3><p>Do the often expensive security controls that are in your environment work as intended? Every budget cycle, we lobby for funding to maintain or purchase security controls to address the ever-changing threats, vulnerabilities and resultant risks that face our businesses.</p><p>Do the security controls work as the vendor presentation said they would in the pre-sales meeting? Have the controls been implemented and configured correctly to prevent or detect attacks? Do you have blind spots in your monitoring and detection capabilities?</p><p>Most pen testers will start “low and slow” and become increasingly more intrusive. What is the threshold where they are detected by each set of controls, if they are detected at all? Can the tester bypass the controls? I once gained access inside a financial institution by getting remote code execution on an unpatched IDS device.</p><p>A noisy pen test should light up the dashboards for your security controls like a Christmas tree.</p><p>Use the pen test as an opportunity to tune the controls or tighten up configurations — e.g. firewall rules — to ensure the testers or, more importantly, the bad guys are detected earlier next time.</p><h3>Tip 11: Test your “watchers” to ensure timely detection.</h3><p>Not informing the Security Operations Center (SOC) or your Managed Security Service Provider (MSSP) of the pen test is an excellent way of testing if those teams are watching the consoles or are drowning in alerts. I’ve seen pen tests complete or others go almost to completion — including breaching the perimeter — without being detected.</p><p>Reviewing how the testers got in and the tools, techniques and processes they used are a great way to help your SOC hone its threat-hunting skills and tune signatures.</p><p>The lessons learned from gaps in detection can also help test and uplift your digital forensics and incident response processes.</p><h3>Summing It Up</h3><p>Pen testing can uncover infrastructure or vulnerabilities you weren’t aware of and help the organization better understand its cybersecurity risks.</p><p>By adopting a targeted approach to penetration testing with well-defined scope and test cases that are company- or objective-specific, you can derive maximum value from your penetration testing spend.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=403d751228fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>