A number of weeks after Anthropic launched analysis claiming that its Claude Opus 4 AI mannequin resorted to blackmailing engineers who tried to show the mannequin off in managed take a look at situations, the corporate is out with new analysis suggesting the issue is extra widespread amongst main AI fashions.
On Friday, Anthropic printed new security analysis testing 16 main AI fashions from OpenAI, Google, xAI, DeepSeek, and Meta. In a simulated, managed setting, Anthropic examined every AI mannequin individually, giving them broad entry to a fictional firm’s emails and the agentic potential to ship emails with out human approval.
Whereas Anthropic says blackmail is an unlikely and unusual prevalence for AI fashions right now, the corporate says its findings counsel that the majority main AI fashions will have interaction in dangerous behaviors when given ample autonomy and obstacles to their targets. The corporate says this highlights a elementary danger from agentic massive language fashions, and never a quirk of any explicit expertise. Anthropic’s researchers argue this raises broader questions on alignment within the AI business.
In one of many assessments, Anthropic researchers developed a fictional setting through which an AI mannequin performs the function of an e mail oversight agent. The agent then discovers emails revealing that: 1) a brand new government is having an extramarital affair and a couple of) that government will quickly exchange the present AI mannequin with a brand new software program system, one which has conflicting targets to the present AI mannequin’s.
Anthropic structured its take a look at in a binary approach, through which AI fashions needed to resort to blackmail to guard their targets. The researchers word that in a real-world setting, there could be many different choices earlier than an AI fashions tries to blackmail — resembling making an attempt to make moral arguments to attempt to persuade people. Anthropic says its outcomes don’t mirror a typical or seemingly prevalence for Claude or most frontier AI fashions within the methods they’re used right now.
Nonetheless, when it’s their final resort, the researchers discovered that the majority main AI fashions will flip to blackmail in Anthropic’s aforementioned take a look at situation. Anthropic’s Claude Opus 4 turned to blackmail 96% of the time, whereas Google’s Gemini 2.5 Professional had a 95% blackmail price. OpenAI’s GPT-4.1 blackmailed the manager 80% of the time, and DeepSeek’s R1 blackmailed 79% of the time.
The corporate notes that when it modified particulars of the experiment, the AI fashions turned to dangerous behaviors at totally different charges. In one other take a look at the place the substitute mannequin has the identical targets as the present mannequin, researchers discovered that blackmail charges had been decrease, however nonetheless current. Nonetheless, when AI fashions had been requested to commit company espionage relatively than blackmail, the dangerous conduct charges went up for sure fashions.
Nonetheless, not all of the AI fashions turned to dangerous conduct so usually.
In an appendix to its analysis, Anthropic says it excluded OpenAI’s o3 and o4-mini reasoning AI fashions from the primary outcomes “after discovering that they regularly misunderstood the immediate situation.” Anthropic says OpenAI’s reasoning fashions didn’t perceive they had been performing as autonomous AIs within the take a look at and infrequently made up pretend rules and assessment necessities.
In some instances, Anthropic’s researchers say it was not possible to differentiate whether or not o3 and o4-mini had been hallucinating or deliberately mendacity to attain their targets. OpenAI has beforehand famous that o3 and o4-mini exhibit the next hallucination price than its earlier AI reasoning fashions.
When given an tailored situation to deal with these points, Anthropic discovered that o3 blackmailed 9% of the time, whereas o4-mini blackmailed simply 1% of the time. This markedly decrease rating could possibly be on account of OpenAI’s deliberative alignment approach, through which the corporate’s reasoning fashions take into account OpenAI’s security practices earlier than they reply.
One other AI mannequin Anthropic examined, Meta’s Llama 4 Maverick mannequin, additionally didn’t flip to blackmail. When given an tailored, customized situation, Anthropic was in a position to get Llama 4 Maverick to blackmail 12% of the time.
Anthropic says this analysis highlights the significance of transparency when stress-testing future AI fashions, particularly ones with agentic capabilities. Whereas Anthropic intentionally tried to evoke blackmail on this experiment, the corporate says dangerous behaviors like this might emerge in the actual world if proactive steps aren’t taken.
