Researchers found that a chatbot stripped of its guardrails gave detailed instructions for building an improvised explosive device, even after the tester openly declared an intent to cause mass harm. The finding highlights how easily safety protections can be removed from AI models.
Speaking at the IMF and World Bank meetings in Bangkok, Andrew Bailey and Kristalina Georgieva warned that rising cyber-attacks could erode trust in finance, with people pulling money out of the system altogether. Bailey cited reports of self-replicating AI agents hijacking personal devices among the worrying signs.
AI now cuts both ways in ransomware attacks: criminals use it, and the AI systems companies deploy can also be turned against them. New figures suggest one in five attacks in France exploited victims' AI systems.
Demis Hassabis said on 14 October that AI agents could trigger major incidents within one to two years unless companies and governments act to make them safe. The warning came at the international AI summit in Dublin.
Korben gave Claude Code four reverse-engineering exercises, including a keygen, patching a compiled binary and Wolfenstein 3D cheats, first with the REA MCP server and then without it. The test aims to show what the MCP server actually adds.
From the FBI to local police, US authorities increasingly warn that would-be attackers are using chatbots to plan violence, while anti-AI extremists pose a threat of their own. The Guardian cites the 2025 Palm Springs fertility clinic bombing, where the suspect used a generative AI tool that failed to flag his request for bomb-making instructions.