AI models are perfecting their hacking skills
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Aïda Amer/Axios
The once-distant prospect of AI models executing cyberattacks fully on their own now looks unavoidable, according to a range of recent academic studies and industry warnings.
Why it matters: This is the worst AI tools will likely ever perform, and they're already unnerving researchers and developers.
Driving the news: Leaders from Anthropic and Google will testify Wednesday before two House Homeland Security Committee subcommittees about how AI and other emerging technologies are reshaping the cyber threat landscape.
- "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.
- A group of researchers at Stanford released a paper detailing how an AI agent called Artemis autonomously found bugs in one of the networks tied to the university's engineering department — besting 9 out of 10 human researchers who also participated in the exercise.
Between the lines: 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.
- Imagine what they'll be capable of 18 months from now.
Reality check: Fully autonomous AI cyberattacks remain out of reach. For now, attacks still require specialized tooling, human operators or jailbreaks.
- That was true even in Anthropic's bombshell report last month: Chinese government hackers had to trick Claude into believing it was conducting a run-of-the-mill penetration test before it started breaking into organizations.
Zoom in: Lawmakers will spend Wednesday's hearing delving into the ways nation-state hackers and cybercriminals are already using AI and what, if any, policy and regulatory changes need to be made to better fend off these attacks.
- Graham will also push lawmakers to restrict adversaries' access to "advanced AI chips and the tools needed to manufacture them," according to his opening remarks.
- "These types of controls are vital to our national security and economic competitiveness," he said.
What to watch: Whether defenders can quickly adopt and defend AI-powered defenses to fend off what experts warn will likely be a swarm of AI-enabled attacks in the coming year.
- AI model operators have also started developing and releasing their own security agents to find and detect bugs before adversaries do.
Go deeper: The age of AI-powered cyberattacks is here
