AI models are perfecting their hacking skills

""We believe this is the first indicator of a future where, despite strong safeguards, AI models may enable threat actors to conduct an unprecedented scale of cyberattacks," Logan Graham, head of Anthropic's AI red team, wrote in his opening testimony, shared first with Axios. "These cyberattacks may become increasingly sophisticated in their nature and scale," he added. Catch up quick: OpenAI warned last week that future frontier models will likely possess cyber capabilities that pose a high risk - significantly lowering the skill and time a user would need to carry out certain types of cyberattacks."

"Researchers at Irregular Labs, which runs security stress tests on frontier models, said they've seen "growing evidence" that AI models are improving in offensive cyber tasks. That includes improvements in reverse engineering, exploit construction, vulnerability chaining and cryptanalysis. Flashback: Just 18 months ago, those models struggled with "basic logic, had limited coding capabilities, and lacked reasoning depth," Irregular Labs noted."

"Reality check: Fully autonomous AI cyberattacks remain out of reach. For now, attacks still require specialized tooling, human operators or jailbreaks."

Leaders from Anthropic and Google will testify before House Homeland Security subcommittees about how AI and emerging technologies are reshaping the cyber threat landscape. Anthropic warned that AI models could enable threat actors to conduct cyberattacks at unprecedented scale and growing sophistication. OpenAI warned that future frontier models may possess cyber capabilities that significantly reduce the skill and time needed to carry out certain attacks. Researchers demonstrated an AI agent that autonomously found network bugs and outperformed most human researchers. Security testers report improvements in reverse engineering, exploit construction, vulnerability chaining, and cryptanalysis. Fully autonomous attacks remain out of reach and still require tooling and human involvement.

Read at Axios