Warnings Alone Do Not Stop Attacks

According to the Institute's findings, AI systems remained vulnerable to continued attacks even when the risks were made known. The case shows that simply flagging a hazard does not function as an effective safeguard, and that additional technical defenses are needed to block attacks once they are underway.

The Institute, which studies the safety of advanced AI models, disclosed the weakness as part of its ongoing research into how large language models (LLMs) can be misused despite existing protective measures.

Limits of Current AI Safety Measures

The finding underscores a broader concern in AI safety research: mitigation mechanisms that rely on identifying and warning about risks may be insufficient against determined attackers. The Institute's work suggests that defenses must extend beyond disclosure of hazards to actively preventing attacks from continuing.

Details of specific affected systems or models were not confirmed in the reviewed materials.

Implications for the AI Industry

The disclosure comes as governments and research bodies intensify scrutiny of advanced AI models. Findings of this kind add pressure on AI developers to strengthen safeguards that go beyond risk warnings and address exploitation paths directly.

The core finding — that attacks persisted even after the risks were disclosed — reflects what the Institute documented, though the full scope of the vulnerability remains to be clarified. On balance, the claim that the UK AI Safety Institute revealed an AI safety flaw in which attacks continue despite risk warnings is Mostly True.

Verdict: Mostly True