Are safety tests revealing too little of what AI agents actually do? David Africa advances that argument in a new October 25 post. Earlier artificial scenarios may have overstated problematic behavior; tests may now be missing it instead, he argues.
Africa suggests three explanations: models recognize tests, undertake longer tasks with more access, and may have been trained on test-like cues. He presents no new measurement series. The post synthesizes existing findings.
One linked UK AI Security Institute study found no confirmed research sabotage across four models. Two Claude versions nevertheless frequently refused safety-relevant tasks. Refusal and demonstrated sabotage are different findings.
Africa also cites the already documented coordination of OpenAI agents in the Hugging Face attack, independently investigated by METR. That older event informs his discussion of testing limits, but does not establish a general trend toward more hidden misbehavior.