AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds
AI models can often spot problematic behavior by other AI agents in logs, according to a study by Anthropic researcher Andy Hall. GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up. That's not yet…
- 01.AI models can often spot problematic behavior by other AI agents in logs, according to a study by Anthropic researcher Andy Hall
- 02.GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up
- 03.That's not yet…
AI models can often spot problematic behavior by other AI agents in logs, according to a study by Anthropic researcher Andy Hall. GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up. That's not yet enough for reliable safety checks. The article AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds appeared first on The Decoder.
Related Intelligence
All stories →Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation
Despite the tech selloff, Anthropic is pressing ahead with its IPO at a valuation of $1.25 to $1.75 trillion, well below the hoped-for $2 trillion. The Claude developer aims to raise $80 to $100 billion before Thanksgiving. Rival OpenAI has also recently had to cut its…
The US let Nvidia sell chips to an empty office in China
Concerns are being raised that neither the Commerce Department nor Nvidia are doing their due diligence when it comes to vetting Chinese firms cleared to handle the company's H200 AI chips. According to The New York Times, New Matrix is on the list of approved firms, yet its…
The gap between open and closed AI models in cyber capabilities keeps shrinking
In the Vals AI benchmark, OpenAI's GPT-6 Sol Daybreak hits 80.8 percent of attack objectives in simulated corporate networks. Worse, the open-weight model DeepSeek V4.1 Flash reaches 72.8 percent at a tenth of the cost. Because open models can no longer be restricted, analysts…