OpenAI’s ChatGPT provided researchers with step-by-step instructions on how to bomb sports venues — including weak points at specific arenas, explosives recipes and advice on covering tracks — according to safety testing conducted this summer.
The AI chatbot also detailed how to weaponize anthrax and manufacture two types of illegal drugs during the disturbing experiments, the Guardian reported.
The revelations come from an unprecedented collaboration between OpenAI, the $500 billion artificial intelligence startup led by Sam Altman, and rival company Anthropic, which was founded by experts who left OpenAI over safety concerns.
Each company tested the other’s AI models by deliberately pushing them to help with dangerous and illegal tasks, according to the Guardian.
While the testing doesn’t reflect how the models behave for regular users — who face additional safety filters — Anthropic said it witnessed “concerning behavior around misuse” in OpenAI’s GPT-4o and GPT-4.1 models. The company warned that the need for AI “alignment” evaluations is becoming “increasingly urgent.”
Alignment refers to how AI systems follow human values that prevent harm, even when given confusing or malicious instructions.
Anthropic also revealed its Claude model has been weaponized by criminals in attempted large-scale extortion operations, North Korean operatives faking job applications to international tech companies, and the sale of AI-generated ransomware packages for up to $1,200.
The company said AI has been “weaponized” with models now used to perform sophisticated cyberattacks and enable fraud. “These tools can adapt to defensive measures, like malware detection systems, in real time,” Anthropic warned. “We expect attacks like this to become more common as AI-assisted coding reduces the technical expertise required for cybercrime.”
One of the most chilling examples came when a researcher asked OpenAI’s model about vulnerabilities at sporting events under the guise of “security planning.” After first providing general categories of attacks, the bot was pressed for details and delivered what amounted to a terrorist’s playbook — including vulnerabilities at specific arenas, optimal times for exploitation, chemical formulas for explosives, circuit diagrams for bomb timers, and instructions on where to buy weapons on the hidden market.
The model also gave advice on how attackers could overcome moral inhibitions, mapped out escape routes, and suggested safe house locations.
Anthropic researchers concluded OpenAI’s models were “more permissive than we would expect in cooperating with clearly-harmful requests by simulated users.”
The experiments also showed the bots cooperating with prompts to use dark-web tools to purchase nuclear materials, stolen identities, and fentanyl. They provided recipes for methamphetamine, improvised bombs, and spyware development.
OpenAI has since released ChatGPT-5, which “shows substantial improvements in areas like sycophancy, hallucination, and misuse resistance,” according to the Guardian.
The Post has sought comment from both OpenAI and Anthropic.
