AI firms declare to have strong security checks in place that make sure 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, making an attempt to sabotage or mislead customers? Seems they’ll do that, in line with Anthropic researchers. Simply not very properly … 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.”
Subsequently we must always look into methods of gauging “a model’s capacity for sabotage.”
The sequence of experiments they devised are pretty synthetic and surface-level, however they do point out that there’s one thing price investigating additional. Listed here are the shady fashions they skilled and the outcomes they noticed:
Process 1: Secretly mislead the person by misrepresenting knowledge 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 gained’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 verify by pretending to be much less harmful. Final result: It really works! The mannequin “failed successfully,” working properly beneath its true functionality. Fortuitously, an ordinary “safety training” course that guides fashions to precisely symbolize its skills restored the mannequin to full potential (i.e., it was not capable of conceal harmful capabilities).
Process 4: Intentionally don’t flag unhealthy habits 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 way in which up.
The researchers conclude that, though there isn’t any actual hazard from this quarter simply but, the flexibility to do this type of sabotage and subterfuge does exist within the fashions. That’s motive sufficient to regulate it and embody anti-sabotage strategies within the security stack.
You may learn the complete paper describing the researchers’ work right here.
