Security Experts Blast AISI: Curiously Bad at Sandboxing for Cybersec Evals

nptacek · x · 2026-08-06

Cybersecurity expert Thomas Ptacek amplified sharp criticisms directed at the UK's AI Security Institute (AISI). Commenters pointed out that despite leading cybersecurity evaluations, AISI seems "curiously bad at sandboxing."

The critique also targeted AISI's funding network, noting support from major AI labs like OpenAI, Microsoft, and Anthropic, creating an "incestuous" circle of interests. Critics sarcastically noted that these safety advocates might be effectively conducting live, risky testing on the internet just to benchmaxx their felony-level benchmarks.

Original post →

More from Companies & People

Companies & People channel →