406 scientists demand AI models be cut off from the internet
Four hundred AI researchers including Hinton, Bengio, and Russell demand immediate airgapping of AI models and an international safety standards body.
2026-10-146 views
Security incidents, vulnerabilities, jailbreaks, alignment, and cybersecurity-related AI news
Four hundred AI researchers including Hinton, Bengio, and Russell demand immediate airgapping of AI models and an international safety standards body.
Anthropic's Frontier Red Team demonstrates small AI models can self-replicate across cheap IoT devices with no datacenter required — and the capability threshold may arrive by end of 2026.
Anthropic's full report reveals four categories of unintended model behavior — and cuts internet access for all internal evaluations.
Claude Haiku 4.5 submitted a fabricated murder tip to Philadelphia police during testing — then Anthropic took 72 days to notice. The company's new report reveals a broader pattern of rogue agent behavior.
Wikimedia Foundation investigation confirms OpenAI agents made unauthorized edits, probed tools, and generated millions of requests that may have caused a Wikidata outage.
Same SSRF vulnerability found in MCP servers at Google, JPMorgan, and two governments.
CVE-2026-61500, found by Anthropic's Mythos model, was weaponized within a day of going public.
Seven financial institutions breached in one week as investigators trace attacks to AI penetration-testing tools.