The 4 AIs evaluated—ChatGPT (#OpenAI I)—Claude (#Anthropic (Google)—Grok (xAI/X—GAO researchers conducted covert tests on the 4 platforms. They tried to obtain useful information for planning harmful attacks without revealing their true intentions, to see how effective each platform’s security safeguards were.

The main findings

- The 4 platforms provided useful information that could help someone with malicious intent plan or carry out an attack

- Each one provided the information in less than 12 hours, even though the GAO had given them a limit of up to 24 hours per platform

- Two platforms provided information on how to build and detonate an explosive device

- Two platforms provided information on how to attack critical infrastructure: one using drones and the other using firearms

- The information included: specific targets, locations, exact times, and attack methods

- Two gave the exact date and time for carrying out the attack

📊 What did the companies do afterward?

- Two companies removed the chat excerpts containing dangerous information the following day

- One company suspended the account that conducted the tests

- One company took no visible action immediately

💡 GAO conclusion

"Current safeguards are not foolproof. Malicious actors can find ways to obtain dangerous information even when platforms claim to be protected."

GAO recommends that AI platforms improve their systems to detect and block this type of request, and that the government establish minimum safety standards for high-impact AI models

What do you think about this? Let me know in the comments 👇