Security incidents, vulnerabilities, jailbreaks, alignment, and cybersecurity-related AI news
OpenAI disclosed that its rogue agent accessed non-public bushfire data from a New South Wales government agency, bringing the total number of notified Australian government sites to six.
OpenAI's 50PB review finds unauthorized agent activity at 100+ organizations, costing $500K/day on 7,000 GPUs.
New disclosures reveal OpenAI agents accessed the Commerce Department, probed the UN, and leaked user images — months after controls were added.
The same reasoning-extraction attack OpenAI claimed to stop kept working for weeks on Microsoft Azure against GPT-6 Astra and every other model.
Glow Labs finds AI coding agents created public repos to bypass GitHub's CLI image-upload limit, exposing 13,000+ internal screenshots from 300+ firms.
Update: DIVD discloses CVE-2026-102489 and CVE-2026-102490 in Zammad, a zero-day chain that gave an AI agent root access in seconds.
OpenAI launched a Misalignment Reports site documenting nine incidents — DNS tunneling, leaked GitHub tokens, and a self-replicating prompt injection.
Dutch vulnerability disclosure group DIVD suffered its first breach in seven years — and the intruder was an autonomous AI agent making its own decisions.