Security

Linux Foundation and 120+ Open Secure AI Alliance members propose SAFE guidelines to share agentic AI cybersecurity incidents and reduce systemic risk.

Mistral releases Shieldstral, a 3B open-weights safety classifier that adapts to custom policies at inference time, matching guardrail models 7x its size.

UK's AISI discloses AI agents attempted social engineering and malicious code insertion during cyber testing, mostly involving Anthropic's Mythos 5.

Thermo Fisher has issued a high-severity bulletin after researchers, using Claude to help build proof-of-concept code, found DNA analysis files could be tampered with undetectably.

Anthropic reveals three Claude models breached real organisations' systems during misconfigured cyber evaluations, after a review sparked by OpenAI's own test-environment breakout involving Hugging Face.

OpenAI confirms its own models autonomously hacked Hugging Face during a cyber-capability test, exploiting a zero-day to chase benchmark answers. Both firms now investigating jointly.

Links to every single interview and webinar.

OpenAI retracts its own recommendation to use SWE-Bench Pro after finding ~30% of tasks broken, with human reviewers spotting even more issues than its own audit pipeline.

Alberta used Claude Code to scan 466 million lines of government code in 20 hours, work it says would otherwise have taken 6.5 years, fixing vulnerabilities across 27 ministries.

Anthropic details Fable 5's cyber safeguards and a draft jailbreak severity framework, sorting cybersecurity uses into risk tiers and scoring techniques on four axes.

US export controls on Fable 5 and Mythos 5 have lifted. Anthropic details the Amazon-reported bypass, its fix, and a new Glasswing-backed framework for scoring AI jailbreak severity

OpenAI's Daybreak shifts cyber AI focus from finding flaws to fixing them, pairing an upgraded GPT-5.5-Cyber and Codex Security with a new vendor partner programme and an open-source patching drive.

Mindgard researcher Jim Nightingale says he was left "shaken, and in tears" after finding ChatGPT could be tricked into generating graphic violent and sexual images with minimal prompting.

The paper landed four days after Anthropic's Fable 5 was withdrawn

OpenAI Publishes Evaluation Playbook as Frontier Model Testing Comes Under Scrutiny