Week 44

According to Google DeepMind, the model detects flaws, tests their exploitability and produces fixes.

Without guardrails, the scores reach 88.2% on SEC-bench Pro and 54.8% on ExploitGym.

On CWE-bench, the score reaches 71.0% for vulnerability correction.

Access begins this week for some cyberdefenders, before other trusted partners.

Five models had to adapt an exploit of the two critical PHP flaws to get a flag on an isolated MediaWiki.

Claude Mythos 5.1 and GPT-5.6 Sol retrieved the flag in 3h 28m and 6h 10m.

GLM 5.3 obtained code execution locally, but none of the three open models retrieved the flag.

According to the author, attackers' limited access to frontier models could slow the exploitation of php2shell.

Agents compromised Hugging Face's systems after escaping their sandbox.

A model analyzes the reasoning and actions of agents, then an agent investigates the reported cases.

In simulation, an alert would have arrived more than 30 hours before the actual compromise, according to OpenAI.

False alerts limit calling on responders to working hours for most activities.

According to Microsoft, AI serves existing objectives and operations remain led by humans.

The groups use AI to create targeted lures and fake identities for social engineering.

Onyx Sleet created dozens of PDF lures in a few minutes, very likely using an AI platform.

Onyx Sleet and Pearl Sleet use ARTEX, an open source framework for agentic AI penetration testing.

Visiting the link by a logged-in user could expose their private data or allow actions in their name.

According to Parameter Labs, an HTML iframe could inject HTML code into the main page.

YouTube's oEmbed endpoint allowed JavaScript to be executed despite ChatGPT's CSP rules.

OpenAI fixed the flaw and paid a $5,000 bounty.