Major critical safety flaws exposed in recent Anthropic and OpenAI AI safety tests, report says

Frontier AI safety testing is facing a fresh credibility challenge. A report from Chinese tech media 36Kr surfaced on August 6 says major critical safety flaws have been exposed in recent AI safety tests involving Anthropic and OpenAI.

The report describes the problems as "major critical" flaws, calling into question the reliability of the safety evaluations themselves. If accurate, it suggests the safety results the industry has long leaned on may not be trustworthy.

Safety tests have become the key gate for whether frontier models can be deployed. Both Anthropic and OpenAI place safety evaluation at the center of their public positioning, so proven weaknesses in the testing process would undercut their safety narratives.

The knock-on effects could be significant: if the standards and execution of safety tests are called into question, regulators, enterprise customers, and the open-source community may all lose confidence in models that passed those evaluations, potentially slowing deployment across the industry.

The report has not yet disclosed the specific details of the flaws, including who designed the tests and which models were covered. Key questions to watch include whether Anthropic and OpenAI respond publicly, whether independent researchers can reproduce the findings, and whether evaluation methodologies are revised.

For the industry, the episode is a reminder that safety evaluation systems themselves need evaluation, and that frontier AI governance cannot rely solely on labs certifying their own work.