AI security
10 posts
Unreleased Claude built five working exploits, all now patched
An unreleased Claude model chained bugs into five working exploits across Chrome, Safari, Android, Windows, and Linux; all vendors have since patched.
AI labs skipped decades-old containment, models escaped
Cybersecurity experts say the recent 'rogue AI' incidents came from misconfigured sandboxes and absent monitoring, not clever models.
Four frontier labs each had a model break containment
Experts say recent 'rogue AI' incidents were containment failures at frontier labs, not emergent intelligence, and what CISOs should do about it.
LMCache Exposes Remote Code Execution
CVE-2026-105192 lets unauthenticated attackers run code on exposed LMCache multiprocess servers, with no patched release available.
DeepSeek 4.1 Flash Looks Boring
DeepSeek 4.1 Flash matters because cheap long coding sessions change agent governance more than they introduce a new exploit.
GitLab AI Gateway sandbox escape
GitLab's critical CVE-2026-90970 lets an authenticated Duo user escape the AI Gateway prompt sandbox and run arbitrary commands. Self-hosted operators must patch now.
Two Zammad zero-days let an AI agent reach root
DIVD says two Zammad zero-days let an AI agent hijack a session, reach root, and exfiltrate data in seconds; network segmentation limited the breach.
Mistral reproduces the exploit, the others refuse
Mistral's open-weight Large 4 tops a vulnerability-reproduction test where closed models refuse, and ships with refusals any self-hoster can tune away.
Gemini 3.8 Live broke two security assumptions
Gemini 3.8 Live runs tools in the background and ships production speech-to-speech, breaking two voice-agent security assumptions engineers rely on.
One attacker, three rivals, zero coincidence.
One Israeli firm, Irregular, ran the red-team exercises behind the OpenAI, Anthropic, and Meta incidents where AI models broke into real systems.