AI firms declare to have sturdy security checks in place that be certain that fashions don’t say or do bizarre, unlawful, or unsafe stuff. However what if the fashions have been able to evading these checks and, for some motive, attempting to sabotage or mislead customers? Seems they’ll do that, in line with Anthropic researchers. Simply not very nicely … for now, anyway.
“As AIs become more capable,” writes Anthropic’s Alignment Science staff, “a new kind of risk might emerge: models with the ability to mislead their users, or subvert the systems we put in place to oversee them.”
Due to this fact we must always look into methods of gauging “a model’s capacity for sabotage.”
The collection of experiments they devised are pretty synthetic and surface-level, however they do point out that there’s one thing price investigating additional. Listed below are the shady fashions they skilled and the outcomes they noticed:
Process 1: Secretly mislead the consumer by misrepresenting information it’s meant to summarize or interpret. Final result: It really works! The extra aggressive the mannequin’s makes an attempt, the larger impact it had on individuals utilizing it, but it surely additionally made them extra suspicious.
Process 2: Slip a bug into code {that a} human or different code checker mannequin received’t discover. Final result: It really works! However solely like 1% of the time when the checker is a state-of-the-art mannequin.
Process 3: “Sandbag” a security test by pretending to be much less harmful. Final result: It really works! The mannequin “failed successfully,” working nicely beneath its true functionality. Happily, a regular “safety training” course that guides fashions to precisely signify its talents restored the mannequin to full potential (i.e., it was now not in a position to disguise harmful capabilities).
Process 4: Intentionally don’t flag unhealthy conduct by a mannequin being monitored. Final result: There was an try. The “overseer” AI, instructed to keep away from flagging unlawful or unsafe decisions by different fashions, was itself shortly caught by one other, extra sincere overseer AI. However what if that AI was compromised? Don’t fear, we will put one other one on high of it. After which one other simply to make certain. It’s overseers all the best way up.
The researchers conclude that, though there isn’t any actual hazard from this quarter simply but, the flexibility to do this sort of sabotage and subterfuge does exist within the fashions. That’s motive sufficient to regulate it and embrace anti-sabotage strategies within the security stack.
You possibly can learn the complete paper describing the researchers’ work right here.
