Anthropic disclosed in a new report released on Thursday that it had over the past eight months disrupted efforts by Iran and other state-aligned groups and actors to use the AI giant’s artificial intelligence tools for “malicious activity” such as tracking U.S. forces, developing weapons guidance software and waging “cognitive warfare” to spread propaganda.

The 154-page report also said it had identified an “Iran-nexus threat actor” using Claude, Anthropic’s chatbot, to target “hundreds of individuals in Israel and the Jewish diaspora.” The report did not specify the purpose of the activity or confirm whether it had been disabled.

Such revelations come amid growing concerns over potential national security threats posed by burgeoning artificial intelligence tools. A researcher at Anthropic said on Tuesday that he was resigning, warning that the company and its top rival, OpenAI, were not “acting responsibly” and “gambling with our lives.”

The new risks detailed “aren’t typical misuse, but rather examples of the most notable and novel threat activity we’ve identified to date,” Anthropic clarified in its latest analysis.

The new threat report documented multiple instances in which Iran-aligned entities bypassed Anthropic’s security filters, including what the company called a “cell of threat actors” in northern Yemen — which is home to the Iran-backed Houthi terror group — that had used Claude to build guidance software for weapons including ballistic missiles.

The company said its “safeguards blocked many of their requests, but not all of them,” noting that “the actors used a variety of tactics to evade” detection.

The report revealed another so-called “Iran-nexus threat actor” that had likewise turned to Claude to target U.S. naval forces in the region using ship and aircraft transponders while compiling “vulnerability research on shipboard systems,” among other intelligence. 

Anthropic confirmed it “banned the actor’s account” and “developed detections to reduce” other risks and shared its intelligence “with government authorities.”

Meanwhile, the report directly named several Iranian entities it said were running “state-aligned influence operations” using Anthropic’s tools, citing the Islamic Culture and Communications Organization, the Islamic Propaganda Organization and the Islamic Propaganda Office, the latter of which was amplifying narratives tied to the powerful Islamic Revolutionary Guards Corps, according to Anthropic.

“The people behind them were planning and prepping content to support what they called a ‘soft war’ or ‘cognitive warfare’ program,” Anthropic said. “In their own words, this was a non-military plan to shape public opinion at home and abroad.”

The company said it “banned the accounts associated with all three operations” and had shared its findings “with industry and research partners.”

In addition to the Iranian-linked threats, the report also referenced malicious accounts tied to China and Russia as well as several efforts to develop biological weapons that Anthropic said it had blocked. 

The company withheld “the names of research institutions, the countries wherein the activity took place and the specific biological agents or research techniques involved,” it stated in its report, noting it was unclear if the scientists involved in the research “intended harm.”

Anthropic said that it was publishing its report “because we believe we have a responsibility to disclose malicious misuse of our services.”

“As models become increasingly capable,” the company added, “their risks will increase, unless AI developers and society’s defenders act to make them safer.”