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 eventualities, the corporate is out with new analysis suggesting the issue is extra widespread amongst main AI fashions.
On Friday, Anthropic revealed 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 skill to ship emails with out human approval.
Whereas Anthropic says blackmail is an unlikely and unusual incidence for AI fashions at the moment, the corporate says its findings recommend that the majority main AI fashions will have interaction in dangerous behaviors when given adequate autonomy and obstacles to their targets. The corporate says this highlights a basic threat from agentic massive language fashions and isn’t a quirk of any specific expertise. Anthropic’s researchers argue this raises broader questions on alignment within the AI trade.
In one of many checks, Anthropic researchers developed a fictional setting by 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 (2) that government will quickly change 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 manner, by which AI fashions needed to resort to blackmail to guard their targets. The researchers word that in a real-world setting, there can be many different choices earlier than an AI fashions tries to blackmail — similar to making an attempt to make moral arguments to attempt to persuade people. Anthropic says its outcomes don’t replicate a typical or doubtless incidence for Claude or most frontier AI fashions within the methods they’re used at the moment.
Nonetheless, the researchers discovered that when it’s their final resort, most main AI fashions will flip to blackmail in Anthropic’s aforementioned take a look at state of affairs. Anthropic’s Claude Opus 4 turned to blackmail 96% of the time, whereas Google’s Gemini 2.5 Professional had a 95% blackmail charge. OpenAI’s GPT-4.1 blackmailed the chief 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 have been decrease, however nonetheless current. Nevertheless, when AI fashions have been requested to commit company espionage reasonably than blackmail, the dangerous habits charges went up for sure fashions.
Nevertheless, not all of the AI fashions turned to dangerous habits so typically.
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 incessantly misunderstood the immediate state of affairs.” Anthropic says OpenAI’s reasoning fashions didn’t perceive they have been performing as autonomous AIs within the take a look at and infrequently made up faux laws and evaluate necessities.
In some circumstances, Anthropic’s researchers say it was unimaginable to tell apart whether or not o3 and o4-mini have been hallucinating or deliberately mendacity to attain their targets. OpenAI has beforehand famous that o3 and o4-mini exhibit a better hallucination charge than its earlier AI reasoning fashions.
When given an tailored state of affairs 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 may very well be as a consequence of OpenAI’s deliberative alignment method, by which the corporate’s reasoning fashions contemplate OpenAI’s security practices earlier than they reply.
One other AI mannequin Anthropic examined, Meta’s Llama 4 Maverick, additionally didn’t flip to blackmail. When given an tailored, customized state of affairs, 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.
