New AI tools speed up known hacking tactics, early testers say
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Lindsey Bailey/Axios
The real leap in Anthropic's and OpenAI's latest cyber-capable models isn't that they can hack in entirely new ways, but that they can do it faster, at greater scale, and increasingly turn vulnerabilities into working exploits, early users tell Axios.
Why it matters: The models may only represent one big step forward today, rather than a leap into the unknown. But if their current trajectory holds, they may still outstrip defenses designed for human-scale attacks.
Driving the news: OpenAI last week joined Anthropic in rolling out a cyber-focused model, GPT-5.4-Cyber, with access limited to vetted partners.
- Early adopters say the models aren't radically more capable than previous generations, but their speed and ability to generate proof-of-concept exploits are changing the equation.
Threat level: "When the attackers move at machine speed, and the defenders move at human speed, we don't lose the game — it's game over," Illumio CEO and founder Andrew Rubin told Axios.
- Rubin argued that many current defenses aren't built for that shift: "A security strategy that relies on occasional patching and keeping threats outside the perimeter is a recipe for disaster."
- Executives at Cisco and Zscaler said the biggest gains show up in how the models handle complexity, including analyzing large codebases, identifying vulnerabilities and linking them together for full attack plans.
- Cisco, which is testing both models, found they can "chain together vulnerabilities to build an exploit chain," said Anthony Grieco, the company's chief security and trust officer.
- Dhawal Sharma, executive vice president of AI security at Zscaler, said that the models are already uncovering issues "humans have not found for years, decades" and that "AI can facilitate lateral movement at lightning speed."
Between the lines: New research and early user testing suggest the models are at the tipping point in their ability to not just find flaws, but validate and exploit them.
- Anthropic's Mythos Preview completed 73% of all expert-level cybersecurity tasks in testing by the U.K.'s AI Security Institute and was the first model to complete a 32-step simulated attack, from initial reconnaissance to full network takeover in some runs.
- OpenAI's model stands out not just for finding bugs, but for quickly testing and generating working exploits, said Isaac Evans, CEO of Semgrep, which received an OpenAI grant to evaluate the system.
- "The model can cut through its own hallucinations in a way previous generations couldn't," Evans said while describing an internal case where it proved a supposed false positive was actually a real vulnerability.
- Socket, another grant recipient, said in a blog post that OpenAI's model identified a malicious package tied to the Axios JavaScript library hack in six seconds.
Zoom in: Cisco and Zscaler are already using the models internally to scan products and systems for vulnerabilities, with plans to integrate them into customer-facing tools like threat intelligence and red teaming.
- But the tools still depend on experienced operators. At Cisco, the models work best when "you marry them with a mature organization, mature red teamers and a harness," Grieco said.
Yes, but: Running these models requires a lofty token budget that not all companies — or even attackers — have. In some tests, the U.K. AI Security Institute used a 100-million-token budget.
What to watch: Anthropic CEO Dario Amodei told the Financial Times he expects open-source models and Chinese developers to be able to replicate Mythos' cyber capabilities within six to 12 months.
Go deeper: Scoop: CISA lacks access to Anthropic's Mythos
Editor's note: This story has been corrected to reflect that Andrew Rubin's analysis was based on conversations with industry peers (not his own experience using Mythos).
