• tl;dr sec
  • Posts
  • [tl;dr sec] #355 - Cyber Chaos Monkey, Offsec Harness Index, Top 10 AWS Security Research of 2026

[tl;dr sec] #355 - Cyber Chaos Monkey, Offsec Harness Index, Top 10 AWS Security Research of 2026

Shopify's AI agent system that continuously attacks their production infra, list of open-source offensive AI agent harnesses, la creme de la cloud 2026

Hey there,

I hope you’ve been doing well!

🎃 San Francisco Halloween

Better late than never - here are some #PeakBayArea Halloween anecdotes.

I didn’t have time to get a costume, so I used one I had from previous years: sexy cop (as you do). But I wanted to spice it up and make it timely, so I used ChatGPT to make me a Wild West Wanted poster for misbehaving AI, so I went as I guess some sort of alignment cop.

First, I went to a house party with improv friends- most didn’t get the poster references.

Then I hung out with some tech friends, who either loved it, who were like, “Why would you bring AI to this, it’s our one day off 💀.”

Other costumes I saw:

  • A couple’s costume: OpenAI and Hugging Face 😅 

  • One friend had a bunch of balloons attached to his torso, and he said he was the “AI bubble.” “Don’t pop me, you’re going to crash the U.S. economy!1!”

  • My friend Maddy Zhang dressed up as PHP2Shell. Too real.

  • Another couple’s costume: Astra and Kimi K3 - I asked Astra if he minded that his girl was distilling him, and he was OK with it.

P.S. No newsletter next week, enjoy Thanksgiving (if you celebrate)!

Sponsor

📣 Ship fast. Fix What's Exploitable.

Plenty of tools generate alerts. Seeing which pipeline risks are exploitable takes context.

The Secure Software Delivery Toolkit is built for teams who want to ship fast without letting scanner noise set the agenda. 

It includes 4 practical resources in one place to benchmark your current SDLC, identify critical security gaps, and build an actionable roadmap for your software delivery lifecycle.

What’s Inside: 

  • DevSecOps Security Best Practices

  • CI/CD Security Best Practices

  • AppSec Best Practices Cheat Sheet

  • State of SDLC Security 2026

👉 Get the Bundle 👈

Nice, these guides sound 🔥 

AppSec

Phrack #73
Some of the articles that stick out to me: a profile on Halvar Flake, “Owning the Truth: Undecidability, unsound kernels, and the forging of proof”, “Reflections on Trusting AI: A Virus Inside an LLM”, “MOLE - A Binary Ninja Plugin for Static Taint Analysis”, “Backdoor attacks against Machine Learning IRL”, “A Journey Into Kernel Exploit Development in the LLM Era”, “Security is failure: Every Patch Is a Guess.”

Java4Shell: Write once, RCE everywhere
Calif’s Quang Luong describes a memory-corruption vulnerability in OpenJDK's Base64 decoder (lol, what?) that allows a remote attacker with no credentials to overwrite heap memory, read it in the style of Heartbleed, and achieve remote code execution. Pre-auth RCE in Keycloak, Atlassian Jira & Confluence, JFrog Artifactory, Jenkins, and more.

“To decode Base64 faster, OpenJDK uses specialized x86-64 CPU instructions to process many bytes at once. We discovered a bug in that optimized machine-level code that lets malformed input overwrite neighboring memory… The corruption itself can be triggered by just 64 A characters followed by a single =". 😅 

💡 Calif is like DJ Khaled these days: another one.

Sponsor

📣 60% of orgs run dependencies published less than 24 hours ago

Datadog analyzed security telemetry from thousands of organizations across AWS, Azure, and Google Cloud. 60% were running dependency versions published less than 24 hours earlier, the window where a compromised package like those in the Shai-Hulud campaign does its damage. Meanwhile, the median cloud service now carries four times as many high and critical vulnerabilities as it did in February.

The 2026 State of Cloud Security report also covers long-lived keys for AI inference services, keyless auth in CI/CD, and AWS data perimeters.perimeters.

Oof more people have to enable dependency cooldowns 😅 I always enjoy Datadog’s state of the union reports, interesting to see what trends they find.

Cloud Security

Tag, You’re It: Authorization Traps in AWS ABAC
DEF CON 2026 talk by CrowdStrike’s Itay Saraf in which he presents a taxonomy of authorization traps in AWS Attribute-Based Access Control (ABAC) implementations, showing that AWS's own managed policies have had these issues. He categorizes authorization traps into three types: alterable guards (where users can modify tags to bypass restrictions), evaluation semantics (like case-sensitivity differences), and action-key incompatibility (where actions don't support specified condition keys).

  1. New AWS IAM Attack Paths and the Framework to Exploit Them by Seth Art

  2. CodeBreach: Infiltrating the AWS Console Supply Chain and Hijacking AWS GitHub Repositories via CodeBuild by Yuval Avrahami and Nir Ohfeld

  3. CloudBashing: Exploiting Free CloudShells for Mining, Networking, Exfil, and Persistence at Scale by Jenko Hwong and Chris Ryan

  4. …and 6 other awesome pieces of research, check it out!

Blue Team

Policy and Disclosure: 2026 Edition
Natalie Silvanovich announces that Google Project Zero is overhauling its vulnerability disclosure policy in response to AI-accelerated threat landscapes, moving from "90 + 30" to a phased "Full X" approach: 90 days starting now, 60 days by February 2027, and 30 days by January 2028. The change eliminates grace periods and variant policies because LLMs enable near-instant patch reverse engineering (making "soak time" obsolete), dramatically increase vulnerability discovery rates (Project Zero now finds several bugs weekly in targets that previously yielded only a few yearly), and allow multiple parties to independently discover the same bugs. The policy removes severity-based exceptions and the requirement to verify fixes before disclosure, streamlining what had become an unmanageable negotiation process as vulnerability volumes surge.

Black October: Lessons From the Downstream Breaches
Friend of the newsletter Ryan McGeehan shares lessons from doing incident response for a number of October 2026 serial platform breaches, where compromised dev and infra platforms led to widespread credential leaks affecting their customers. He recommends three key defenses:

  1. “Full stack” network lockdown - Implementing network allow lists across the full stack (VPNs, security groups, and platform-specific IP restrictions for GitHub, AWS, Okta, etc.) which would have prevented 90%+ of the incidents he observed.

  2. Reduce and manage secrets - Reducing secrets by using TruffleHog to inventory exposed credentials, replacing them with Workload Identity Federation or role-based access, and storing remaining secrets in managers like HashiCorp Vault or AWS Secrets Manager.

  3. Establish forensic-ready logging - With centralized collection, consistent formatting, and tamper-resistant retention to enable rapid incident response and customer disclosure.

💡 Highly recommend the rest of Ryan’s blog, it is 👌 

AI + Security

Quicklinks

  • AI Hacking Incidents - An attempt to track cyberattacks where attackers used AI, from published reports. Currently 286 AI-assisted cyberattacks from January 2023 to November 2026. Created by my bud Daniel Miessler.

  • Cheat Sheet: Claude Code Security Best Practices - Claude Code and other AI coding assistants are changing how fast code ships — but they're also changing what security teams need to watch for. This practical cheat sheet breaks down the real risk surfaces AI-assisted development introduces, and gives cloud security, AppSec, and platform engineers tactical steps to close the gaps.*

  • Offsec Harness Index - An attempt to track the open-source offensive AI agent harnesses we can verify on GitHub, and to show which ones attackers are known to have run in real operations. By Daniel Miessler.

*Sponsored


Inside the loop: Five AI attack frameworks that never ask permission
Elastic’s Mika Ayenson analyzes five agentic offensive frameworks (ARTEX, CyberStrikeAI, Cairn, PentAGI, and HexStrike) that use LLMs to autonomously plan and execute attacks in 2-25 second bursts, finding that none enforce authorization in code by default and three explicitly tell their models they're pre-authorized. The frameworks leave traditional telemetry (processes, web requests, API calls, sign-ins) that defenders already collect, but their speed demands pre-selecting which detections trigger automated containment, and their ability to switch infrastructure when blocked means behavior-based detections outlast those keyed on source IPs or User-Agents.

Mika provides an ES|QL hunt that detects tool-family bursts by counting distinct client types (curl, Python, AWS CLI, browsers) per source IP in one-minute windows, and recommends combining any IP blocks with credential revocation, measuring tempo per identity, scoping assumed roles tightly, and removing secrets from accessible locations.

💡 Some nice diagrams, and overall good overview, I like it 👍️ 

Cyber Chaos Monkey: Staying ahead of security threats with AI agents
Shopify’s Dan Scofield describes Cyber Chaos Monkey (CCM), an AI agent system that continuously attacks their production infrastructure using live credentials, source code, threat intelligence, past reports, and infrastructure records to find vulnerabilities before external attackers. The system uses a two-component architecture: Discovery (which uses frontier cyber models to identify potential vulnerabilities) and Sandbox (where agents exploit bugs via live requests, log analysis, and fuzzing).

CCM ensures safety through containerization, egress controls, credential proxies, and a "two-person integrity" approach where a separately-aligned judge agent monitors and can instantly kill the executor agent if it violates Rules of Engagement. CCM has executed tens of thousands of offensive engagements and identified dozens of impactful findings without causing outages.

Misc

✉️ Wrapping Up

Have questions, comments, or feedback? Just reply directly, I’d love to hear from you.

If you find this newsletter useful and know other people who would too, I'd really appreciate if you'd forward it to them 🙏

Thanks for reading!

Cheers,
Clint

P.S. Feel free to connect with me on LinkedIn 👋