Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5, two frontier AI models, cleared autonomous cyber tasks at a pace that left existing capability forecasts trailing, researchers reported Wednesday.
Frontier Models Outpace Trend
Claude Mythos Preview and GPT-5.5 outperformed the doubling trend that the United Kingdom's AI Security Institute had been tracking since late 2024.
AISI had estimated earlier this year that the 80% reliability cyber time horizon was doubling roughly every five months, down from an eight-month figure in November 2025. A newer checkpoint of Mythos Preview solved "The Last Ones," a 32-step simulated corporate network attack, in 6 of 10 attempts, and finished "Cooling Tower" in 3 of 10. GPT-5.5 cleared "The Last Ones" in 3 of 10 attempts.
It was the first time any model had completed both AISI ranges.
Also Read: Ripple's Schwartz Says Bitcoin's Mining Model Is The Flaw XRP Avoided
Palo Alto Vulnpocalypse Warning
Palo Alto Networks reported similar conclusions through its own testing.
The firm scanned more than 130 products over the past month and uncovered 75 legitimate vulnerabilities, more than seven times its typical monthly count, all now patched.
Lee Klarich, the company's tech chief, said organizations have a narrow window before adversaries gain similar capabilities.
He estimated a "narrow three-to-five-month window for organizations to outpace the adversary."
The AISI cautioned that its sample remains small and the hardest tasks have limited human comparison data. Even so, the institute said dropping any single model from the analysis shifts the doubling estimate by less than a month. METR, a nonprofit that tracks AI on software tasks, arrived at a near-identical figure of roughly four months.
Capability Curve Steepens
Anthropic limited the early Mythos rollout last month to a select group including Palo Alto Networks, CrowdStrike, Amazon, Apple and JPMorgan.
OpenAI followed with its GPT-5.5-Cyber model and its Daybreak cyber initiative.
The pace of change has steepened sharply over the past 18 months. AISI compressed its projected doubling period from eight months to 4.7 months in February 2026, then compressed it again after this latest round. The recalculated figure now sits closer to four months, mirroring METR's reading on software engineering tasks.
Read Next: Coinbase Opens $100K USDC Loans Against Solana Token Holdings