<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
<link href="https://ai.code-maven.com/atom" rel="self" />
<title>AI Maven</title>
<id>https://ai.code-maven.com</id>
<updated>2026-11-16T14:30:01</updated>

  <entry>
    <title>Synthetic Scars: Giving Autonomous AI Agents the Battle-Tested Memory of a Senior Engineer</title>
    <summary type="html"><![CDATA[]]></summary>
    <updated>2026-11-16T14:30:01Z</updated>
    <pubDate>2026-11-16T14:30:01Z</pubDate>
    <link rel="alternate" type="text/html" href="https://ai.code-maven.com/synthetic-scars-giving-autonomous-ai-agents-the-battle-tested-memory-of-a-senior-engineer" />
    <id>https://ai.code-maven.com/synthetic-scars-giving-autonomous-ai-agents-the-battle-tested-memory-of-a-senior-engineer</id>
    <content type="html"><![CDATA[<h2 class="title is-4">Description</h2>
<p>Large Language Models (LLMs) can write code at blistering speed, but any developer who has used them on real-world codebases has noticed the same frustrating paradox: the code looks clean, compiles cleanly, and reads convincingly - yet it walks directly into classic production traps. From async lifecycle leaks and naive mock coercion to race conditions and edge-case regressions, AI assistants routinely generate plausible code that breaks in production.</p>
<p>Why? Because models optimize for next-token probability, not production survivability. LLMs have ingested millions of lines of code, but they have never been paged at 3:00 AM. They don’t have <strong>scars</strong>.</p>
<p>In this talk, veteran systems architect and open-source author Randal L. Schwartz introduces the <strong>Synthetic Scar Architecture</strong> - a practical framework that turns production failures, subtle regressions, and debugging post-mortems into permanent, self-enforcing defenses for autonomous coding agents.</p>
<p>Rather than relying on naive prompt engineering or wishful system instructions (which agents routinely ignore or &quot;rationalize away&quot;), the Scar framework structures engineering wisdom into a three-part reflex:</p>
<ol>
<li><strong>The Wound</strong>: The exact runtime failure, leak, or silent regression encountered.</li>
<li><strong>The Trap</strong>: The statistically probable, seductive code pattern the LLM naturally wants to generate.</li>
<li><strong>The Permanent Reflex</strong>: The non-negotiable invariant and test-driven verification gate that prevents the mistake from ever recurring.</li>
</ol>
<p>Randal will demonstrate how this closed-loop consolidation turns transient developer corrections into durable agent memory, how to pair it with rigorous Test-Driven Development (TDD) barriers, and how engineering teams can apply these principles today - whether working with terminal-based autonomous agents or everyday IDE pair-programmers - to eliminate repeat regressions once and for all.</p>
<h2 class="title is-4">Bio</h2>
<p><a href="https://www.linkedin.com/in/randalschwartz/">Randal Schwartz</a> is a self-taught programmer, writer, trainer, and new media host with a passion for technology and creative pursuits.</p>
<p>Throughout his career, Randal has honed his skills in various programming languages, including Perl, Dart, and Flutter, and has become a recognized expert in the field. Notably, he has authored several influential books on Perl programming, including &quot;Programming Perl,&quot; &quot;Learning Perl,&quot; and &quot;Effective Perl Programming.&quot;</p>
<p>He is currently recognized as a Google Developer Expert in the areas of Dart and Flutter (one of 10 in the United States and 150 in the world).</p>
<p>Randal's professional journey has taken him through diverse roles, from software developer and system administrator to consultant and technical writer. He has contributed his expertise to numerous organizations, including Stonehenge Consulting Services, Inc., O'Reilly &amp; Associates, and TWiT.tv, where he hosted the popular show &quot;FLOSS Weekly.&quot;</p>
<p>Beyond his technical prowess, Randal is also a gifted communicator and educator. He has lectured at conferences, provided technical training, and shared his insights through magazine articles and online platforms.</p>
<p>Randal's unique blend of technical expertise, writing talent, and engaging personality has made him a sought-after speaker, author, and consultant in the tech industry.</p>
<h2 class="title is-4">Length</h2>
<p>60 minutes</p>
<p><a class="button is-primary" href="https://luma.com/8chl540w">register</a></p>
]]></content>
    <author>
      <name></name>
    </author>
  </entry>

  <entry>
    <title>Securing the AI Harness: Security in the Agentic Lifecycle</title>
    <summary type="html"><![CDATA[]]></summary>
    <updated>2026-11-16T12:30:01Z</updated>
    <pubDate>2026-11-16T12:30:01Z</pubDate>
    <link rel="alternate" type="text/html" href="https://ai.code-maven.com/securing-the-ai-harness-security-in-the-agentic-lifecycle" />
    <id>https://ai.code-maven.com/securing-the-ai-harness-security-in-the-agentic-lifecycle</id>
    <content type="html"><![CDATA[<p>As AI agents become increasingly integrated into software development, we are not only engineering models, but the harnesses around them - their tools, context, credentials, and interactions with external systems. This creates new security surfaces where sensitive data can unintentionally flow from trusted environments into model context, telemetry, logs, or external services. This lecture explores the security vulnerabilities that emerge within harness engineering, how to reason about sensitive-data flow across the agentic lifecycle, and practical best practices for reducing these risks. It will also introduce Weir, an open-source tool for deterministically testing and monitoring credential flows in AI agents.</p>
<h2 class="title is-4">Bio</h2>
<p><a href="https://www.linkedin.com/in/ido-goldstein-9ba87b152/">Ido Goldstein</a></p>
<p>Software Engineer from big tech (Meta), Cyber startups (Token, Ironscales) and IDF. I innovate security in Agentic Development Lifecycle.</p>
<h2 class="title is-4">Length</h2>
<p>40 minutes</p>
<p><a class="button is-primary" href="https://luma.com/5hwm0xf2">register</a></p>
]]></content>
    <author>
      <name></name>
    </author>
  </entry>

  <entry>
    <title>AI Maven</title>
    <summary type="html"><![CDATA[]]></summary>
    <updated>2026-11-16T10:00:03Z</updated>
    <pubDate>2026-11-16T10:00:03Z</pubDate>
    <link rel="alternate" type="text/html" href="https://ai.code-maven.com/" />
    <id>https://ai.code-maven.com/</id>
    <content type="html"><![CDATA[<h2 class="title is-4">Welcome to the AI Maven!</h2>
<ul>
<li>
<p>See the <a href="https://luma.com/ai-maven">AI Maven</a> calendar.</p>
</li>
</ul>
]]></content>
    <author>
      <name></name>
    </author>
  </entry>

  <entry>
    <title>Poking AI, Finding a 20 years old RCE</title>
    <summary type="html"><![CDATA[]]></summary>
    <updated>2026-11-08T19:30:01Z</updated>
    <pubDate>2026-11-08T19:30:01Z</pubDate>
    <link rel="alternate" type="text/html" href="https://ai.code-maven.com/poking-ai-finding-a-20-years-old-rce" />
    <id>https://ai.code-maven.com/poking-ai-finding-a-20-years-old-rce</id>
    <content type="html"><![CDATA[<p>What happens when you spend half an hour testing an AI model's safety guardrails on offensive security tasks? Sometimes, you accidentally find a critical vulnerability in the backbone of the Perl ecosystem.</p>
<p>In this talk, I'll share the story of how a casual 30-minute experiment testing DeepSeek's rejection rates turned into the discovery of an RCE in PAUSE (the Perl Authors Upload Server) that had been lurking in the codebase for over two decades.</p>
<p>We'll walk through how a quick exploration led to Perl's infamous 2-argument <code>open()</code> footgun, how an uploaded archive filename could trigger code execution, and how combining that with a local privilege escalation could have allowed an attacker to poison CPAN package mirrors worldwide. We'll also chat about the role of LLMs in vulnerability hunting today—where they excel, where they stumble, and how casual curiosity can still yield major security findings.</p>
<h2 class="title is-4">Bio</h2>
<p><a href="https://www.linkedin.com/in/imre-rad-2358749b/">Imre Rad</a> has spent over a decade in IT security across offensive operations, defensive engineering, and software development. Having coded heavily in Perl back when it was popular, they bring a strong historical lens to security research alongside extensive experience in vulnerability programs—both as an active bug bounty hunter and as a program lead for a major enterprise VRP. Today, they handle security for a fast-growing AI startup, while spending their little spare private capacity on security again.</p>
<h2 class="title is-4">Length</h2>
<p>30 min</p>
<h2 class="title is-4">Other titles:</h2>
<ul>
<li>Breaking CPAN: Inside the PAUSE RCE</li>
</ul>
<p><a class="button is-primary" href="https://luma.com/8h4kzkwo">register</a></p>
]]></content>
    <author>
      <name></name>
    </author>
  </entry>

  <entry>
    <title>Mechanistic interpretability for AI Security</title>
    <summary type="html"><![CDATA[]]></summary>
    <updated>2026-11-05T10:30:01Z</updated>
    <pubDate>2026-11-05T10:30:01Z</pubDate>
    <link rel="alternate" type="text/html" href="https://ai.code-maven.com/mechanistic-interpretability-for-ai-security" />
    <id>https://ai.code-maven.com/mechanistic-interpretability-for-ai-security</id>
    <content type="html"><![CDATA[<p>Large language models are usually tested from the outside, prompt by prompt, and judged by what they say. Mechanistic interpretability takes the opposite stance: it treats a model as an engineered object whose internal computation can be read, decomposed and causally edited. Linear probes recover concepts from activations, steering vectors add or subtract a direction, activation patching swaps one component’s output for another’s, and sparse decompositions split a layer into named features. The central finding of recent years is that behaviours such as refusal, deception and task intent occupy low-dimensional, approximately linear subspaces that are stable across prompts.
For security this cuts both ways. A guardrail that lives in a thin slice of activation space can be measured, and a guardrail that can be measured can be removed, rotated, poisoned or watched. The same instruments therefore equip a white-box red team and a white-box blue team. Offensively, they quantify the work factor of a model's intrinsic safeguards: how much weight noise, how small an activation push, how shallow an ablation suffices before the model complies with what it would otherwise refuse. Defensively, the same signals become runtime detectors over hidden states and model-diffing tools that reveal what a fine-tune or supply-chain modification changed inside the weights.</p>
<h2 class="title is-4">Bio</h2>
<p><a href="https://www.linkedin.com/in/natan-katz-2936425/">Natan Katz</a> is Co-founder and Chief Scientist of LuminAI Security, an Israeli startup building runtime security for open-weight language models. LuminAI’s approach is grounded in mechanistic interpretability: instead of filtering prompts and outputs, it monitors the model's internal beahvor as it runs and detects malicious activities and cyber threats from inside the model.</p>
<p>Before LuminAI, Natan led AI technology at Avanan/Check Point, developing ML systems for detecting malicious code, HTML phishing, and network threats. In Web3 security, he has researched ML-based detection of malicious Ethereum transactions and smart-contract vulnerabilities, including LLM-based analysis with CodeLlama and gas-driven detection methods. Earlier, as Principal AI Researcher at NICE, he built AI for fraud-ring detection and voice biometrics. He holds more than 15 patents in applied AI, including graph-neural-network systems for fraud and malicious-code detection, and has represented Israel on an international standardization committee for quantum computing.</p>
<p>Natan trained as a mathematician (Hebrew University, MSc Weizmann Institute) and was a guest scientist in non-linear dynamics at Weizmann and Goethe University Frankfurt.</p>
<h2 class="title is-4">Length</h2>
<p>40 minutes</p>
<p><a class="button is-primary" href="https://luma.com/hw2bkl7x">register</a></p>
]]></content>
    <author>
      <name></name>
    </author>
  </entry>

</feed>

