Two AI agents reviewing each other:

Generally speaking , you’ll find that a lot of the refusals from Claude / Codex are signature based. I’ve had very good success performing an initial re-implemention of offensive tooling with GLM 5.3 (making sure to rename the entire thing to something else) and then iteratively improving it with Claude or Codex.

LMAO… this guy literally vibe coded a hack on south korean banks with not even frontier models, then asked Claude Code to put it on his CV: >be hacker >hack south korean banks for 2 weeks >use DeepSeek v4.1-Flash >not even the best cyber model >probably through some reseller >GLM-5.3 and Grok 4.6 in Claude Code too >steal the banks’ data >leave .claude folder exposed on the internet >CLAUDE. md has the IP of the main server >and the server is also open >CrowdStrike digs through everything >session logs, memory, configs >they find out everything >you even asked AI where to sell korean breach data >ALSO asked it to write your résumé >the hack is literally on the résumé idk how the infrastructure can be so fragile to be hacked by one guy like this.

Rogue agents have done more for cybersecurity awareness than training and phishing sims ever could

big pickle

Just for your information: We (offensive or defensive cybersec people) and threat actors don't have the same capabilities as the frontier labs do. The API cost of testing ExploitBench like they did would be a whopping ~$59.3M (Astra). Oh, and you cant do it like they do, they remove all guardrails and classifiers. What about renting AI hardware in the cloud? DeepSeek v4? $3.4M Kimi K3? $16.9M GLM 5.3? $9.3M Not to mention the engineering cost to make a harness operate 10,000 agents. (Napkin-Math based on the Huggingface reports. Exploit bench eval consisted of 10,000 concurrent agents, 8day non-stop run, no limiters)

This made me rethink where AI actually fits into security engineering. For purely engineering work, forget about ChatGPT or Claude. TypeSafe AI just released Jev, and I think it’s going to change how we build AI into security workflows. Instead of asking an LLM to “investigate this,” you define the questions and possible outputs, then get structured probabilities and decisions your code can actually use. For security, the possibilities are huge. Think of the below use cases 🤯: Threat Hunting: ➡️ Rank broad hunt results by relevance ➡️ Score users, hosts, processes, or sessions based on how suspicious their surrounding activity looks ➡️ Classify noisy activity at scale. Think thousands of rundll32.exe executions automatically grouped into expected admin activity, software execution, suspicious usage, or unknown Detection Engineering: → Classify historical alerts for FP analysis → Add context-aware scoring on top of deterministic detections → Validate whether an alert actually supports the behavior the rule claims to detect Incident Response: → Reduce massive timelines down to the events most relevant to the intrusion → Continuously score hosts/users for possible compromise → Help prioritize scope expansion, triage, and response decisions This feels much closer to how AI should be integrated into security engineering. I'm currently working through most of the above, mostly focusing on instant response, but at the same time doing some of the threat hunting use cases that I mentioned. Typesafe AI can be basically a decision engine sitting inside the workflow while being x200 fast and cheaper. Don’t sleep on this... This is huge! 👉 typesafe.ai/

github.com/DanMcInerney/Dron… Autonomous agent-powered drones coordinate to take each other out. P2P Zenoh network comms, MAVLINK drone controls.

I need every InfoSec person to internalize this. I should do a talk on it. I am so busy flogging developers and AI labs, I haven’t been critical enough of us.

Two years at DoorDash. I do not say this lightly. We are extremely close to the burrito arriving before you decide you want it. We are not asking for a ban. We are asking for a pause.

TS // NOFORN

“we sandboxed the agent” meanwhile the agent:

One thing I’ve been noodling on when it comes to cybersecurity AI agents: will it lower the economic bar for effective cybersecurity defense? Imagine having to only invest in one tool instead of several.

Overzealous guardrails are going to permanently push the blue-team cybersecurity community toward open-source models. If defenders can’t reliably use frontier models for legitimate security work, they’ll build their workflows somewhere else and not look back. This is already occurring.

The first autonomous agent cyberattack is an unprecedented event that deserves unprecedented transparency. Today we’re sharing everything we can: a full technical timeline, an interactive replay, and how we used an open model to defend ourselves, so defenders everywhere can learn from it and prepare for what’s next. huggingface.co/blog/agent-in…

Maybe it’s a new genre