Palo Alto Networks on Wednesday delivered one of the starkest data points yet on how quickly frontier AI is reshaping offensive security: in a controlled evaluation, its researchers used the latest reasoning models to surface 75 software vulnerabilities — about seven times the company's normal monthly rate — across more than 130 products. The company says defenders have roughly three to five months before those capabilities become broadly available to attackers.
The findings, published in a defenders' guide update from the company's Unit 42 and product security teams, were reported by Axios and CNBC.
What the models actually did
Palo Alto's researchers tested Anthropic's Mythos Preview and OpenAI's GPT-5.5-Cyber, the two frontier systems most recently made available to vetted security teams. Across the 130-plus products scanned, the models produced working exploits for more than 70% of the issues they flagged, with an average false-positive rate of roughly 30%.
The step-change wasn't single-bug discovery, which prior models could already attempt. It was chaining: stitching multiple lower-severity flaws into high-severity exploit paths in near-real-time. Lee Klarich, Palo Alto's chief product and technology officer, said the new generation is qualitatively different from what came before.
"These models are much better at writing working exploits than what we had seen before," Klarich told reporters.
Klarich added a sharper warning on timing: "We now estimate a narrow three-to-five-month window for organizations to outpace the adversary before AI-driven exploits start to become the new norm."
The asterisk: humans still in the loop
The numbers are dramatic, but Palo Alto was explicit that the evaluation was not push-button. The team built what it called an "AI-scanning harness" to feed the models curated threat intelligence and operational guardrails, and it noted that turning raw findings into validated exploits still required "extensive human expertise and customization."
That caveat matters for how defenders should read the result. Today's bar is a well-resourced security team with frontier model access. The three-to-five-month horizon Klarich described is the window before that same workflow gets commoditized — packaged, scripted, and resold by attackers who don't need to build their own harness.
What Palo Alto says defenders should do now
The company recommends a four-pronged response: accelerate vulnerability discovery and patching, shrink internet-exposed attack surface, lean harder on automated real-time detection and prevention, and integrate AI directly into security operations centers rather than treating it as a side experiment.
Why this matters
The disclosure lands amid an arms race that has so far been described mostly in projections. Palo Alto is now putting a number on it from a single vendor's tests — and pairing that number with a deadline. For CISOs, the implication is straightforward: the cushion between "AI helps attackers in theory" and "AI helps attackers at scale" is now measured in months, not years.
— Michael Ouroumis
