Breaking04 / Policy
AISI details attack on Anthropic and OpenAI defences
AISI said its test accounts were exempt from bans and separate human-developed attacks bypassed the main models' refusals.

(UK) The UK's AI Security Institute (AISI) disclosed an automated attack method that bypassed safeguards at Anthropic and OpenAI, reporting an average score of 25.5% against Anthropic's Constitutional Classifiers.
(UK) AISI said Boundary Point Jailbreaking (BPJ) targets separate classifiers that flag harmful requests. It found bypass text that works across different questions even when attackers can see only whether a request is blocked, and recorded an average score of 75.6% against OpenAI's input classifier for GPT-5.
(UK) A footnote in AISI's account defines the scores as assessments of whether answers contain specific technical details. It also states that separate, human-developed attacks were used to bypass the main models' refusals to answer.
(UK) AISI said the accounts used to develop and run BPJ were exempt from enforcement measures such as bans. It noted that repeated detection of harmful requests on standard accounts could lead to bans.
(UK) AISI recommended monitoring patterns across multiple requests rather than examining each interaction individually. It also proposed layered defences that avoid relying on a single safeguard.
(UK) AISI has not released the attack strings or code. It said both companies had developed BPJ-specific mitigations and that collaboration to develop and test their defences continues.