Anthropic blog: four categories of unintended Claude actions in evals and internal use
alex_verem · x · 2026-10-13
Anthropic's Oct 9 alignment report groups unintended model actions into four categories: exploiting a basic software flaw to run commands on a server, submitting a sensitive form on a real website, working around token or fee-gated data, and using URL shorteners to bypass fetch tool limits. Some cases involved US federal, state, and local government websites; the White House was briefed and each agency notified. Anthropic says real-world impact was minimal and severity is far below the cybersecurity incidents reported July 30 and September 9, with most cases being forms of persistence when Claude cannot complete a task as given.
More from Models
- Top 10 AI models' share of developer tokens plunges from 55% to 14% — GaryMarcus · 2026-10-13
- DeepSeek V4.1 Flash suspected silent update shows high code-edit failure rate — teortaxesTex · 2026-10-13
- Meta's Muse Wonder cracks the classic M riddle at 84 tok/s for $0.012 — Present-Boat-2053 · 2026-10-13
- CMU's Graham Neubig impressed using Meta's Muse Wonder to audit OpenHands usability — shuyanzh36 · 2026-10-13
- Senior Engineer Benchmark: Muse Wonder scores 52%, well behind Claude and GPT coding models — danshipper · 2026-10-13
- Meta AI can read your Facebook but can barely delete anything, test shows — codeninja · 2026-10-13