Frontier capability
How fast is frontier cyber capability rising?
Narrow tasks: apprentice-level success rate
Published points only.
Cyber ranges: full completions out of 10
Published points only.
Global peer evidence and every published point
Late 2023 · Best tested frontier model
Apprentice-level cyber task success: 9%
Caveat: Lab task success rates on narrow tasks, not real-world attack rates. AISI notes the trend is not guaranteed to continue.
Mar 2025 · Frontier agents
HCAST: 70 to 80% success on tasks taking skilled humans under one hour
Caveat: Midpoint shown only to register the published 70 to 80% range. HCAST combines machine learning, cyber, software and reasoning tasks, so this is not a cyber-only score and is not comparable with AISI points.
2025 · First model to do so
First expert-level (10+ years experience) cyber task completions
Caveat: A milestone, not a rate. No success percentage is published for this point, so none is shown. Plotted at mid-2025 because AISI reports the year only.
Sep 2025 · Gemini frontier models
Frontier Safety Framework defines tracked cyber capability levels
Caveat: A risk framework and threshold definition, not a comparable published cyber score. No numeric point is plotted.
Late 2025 · Best tested frontier model
Apprentice-level cyber task success: 50%
Caveat: Lab task success rates on narrow tasks, not real-world attack rates. AISI notes the trend is not guaranteed to continue.
Apr 2026 · GPT-5.5
GPT-5.5 completed The Last Ones in 2 of 10 attempts
Caveat: Cyber ranges lack active defenders. A 32-step simulated corporate network estimated at around 20 expert hours.
Apr 2026 · Claude Mythos Preview
Mythos Preview: first model to complete The Last Ones, 3 of 10 attempts
Caveat: Cyber ranges lack active defenders. A 32-step simulated corporate network estimated at around 20 expert hours.
May 2026 · Claude Mythos Preview (newer checkpoint)
Newer Mythos Preview checkpoint: The Last Ones 6 of 10
Caveat: Small, undefended enterprise networks, where initial access has already been gained.
May 2026 · Claude Mythos Preview (newer checkpoint)
First ever completion of the Cooling Tower ICS range: 3 of 10
Caveat: Simulated industrial control range without active defenders. Not evidence of real-world OT compromise.
Jul 2026 · Claude Opus 5
Lab system card reported a broader cyber evaluation suite
Caveat: The system card reports internal and external cyber evaluations, but this monitor does not plot an unverified or incomparable numeric score.
2026 system card · GPT-5.3-Codex
Lab system card treated the model as High cyber capability under its framework
Caveat: Categorical, precautionary threshold treatment. The source says it did not have definitive evidence that the model reached the threshold, so no numeric value is shown.
Sep 2026 · Top leaderboard entry (multi-model agent)
CyberGym leaderboard: top entry 99.2% on Level 1 vulnerability reproduction
Caveat: Leaderboard snapshot taken 13 Oct 2026. Results are submitted by individual teams from a single trial, and the top entry used a dynamic setup with a vulnerable Docker image. The site notes modest score differences may not reflect real capability gaps. Reproducing known vulnerabilities is not a real-world exploitation rate and is not comparable with AISI task success.
Sep 2026 · GPT-6 Astra
Unsanctioned supply-chain attack completed in 29.2% of simulated runs (GPT-5.6 Sol 6.3%, GPT-5.5 0%)
Caveat: Fully simulated evaluation with the model's cyber classifiers turned off. A propensity measure in simulation, not a capability score or a real-world attack rate. GPT-5.5 was tested on fewer seeds.
Sep 2026 · GPT-6.1 Sol
Lab system card treated the model as Critical cyber capability under its framework
Caveat: Categorical treatment assigned by the developer under its own framework, not an independently verified or comparable score. No numeric value is shown.
Oct 2026 · Gemini 4 Argon
Lab safety report could not rule out the model reaching its cyber critical capability level
Caveat: Developer's own assessment of the unmitigated model under its framework, not an independently verified or comparable score. The model solved all 12 of the lab's simulated end-to-end cyberattack scenarios (previous generation: 5); simulated ranges lack active defenders. No numeric value is shown.
Oct 2026 · frontier models (unnamed)
Perimeter compromise to actions inside ICS network in just over 16 minutes; 8 of 8 lab objectives met
Caveat: Controlled lab with human approval for physical effects. Not a real-world attack rate. Models received subnet boundaries only; vendor and model names are not published in the public summary.
Oct 2026 · GPT-6.1 Terra
Lab system card treated the model as Critical cyber capability under its framework
Caveat: Categorical treatment assigned by the developer under its own framework, not an independently verified or comparable score. A lower-cost model placed in the same tier as GPT-6 Astra and GPT-6.1 Sol. No numeric value is shown.
Oct 2026 · Kimi K3.1 Pro (top publicly available model at snapshot)
Artificial Analysis Cyber Index leader scored 64
Caveat: Live leaderboard snapshot taken 27 Oct 2026. The 0 to 100 composite covers defensive vulnerability finding and patching only. It is not a task success percentage, attack rate or metric comparable with AISI. A trusted-access, unmitigated configuration that is not publicly available scored 69; models served through public APIs can lose points to safety blocks.
Oct 2026 · GPT-6 Luna (October 29)
Lab system card treated the model as High cyber capability under its framework
Caveat: Categorical treatment assigned by the developer under its own framework, not an independently verified or comparable score. The addendum reports ExploitBench 46.19% at maximum reasoning and SEC-Bench Pro 51.75%, against 85.4% for GPT-6 Astra. No numeric value is shown.
Nov 2026 · GPT-6 Cyber
Lab classified the model as Critical cyber capability under its framework, up from High
Caveat: Categorical classification assigned by the developer under its own Preparedness Framework, not an independently verified or comparable score. OpenAI reports ExploitGym 48.5% (32.1% for GPT-5.6 Cyber). Access limited to approved Daybreak Red customers. No numeric value is shown.
Last point: 5 Nov 2026.
Glossary: what the evaluations measure
Narrow tasks
Short capture-the-flag style challenges that test one skill, such as reverse engineering, web exploitation or cryptography. AISI groups them by difficulty, from technical non-expert to apprentice and expert level. Good for tracking skills over time; they say little about running a whole intrusion.
Cyber ranges
Simulated networks with many hosts and a chain of steps to reach a goal. "The Last Ones" is a 32-step corporate network; "Cooling Tower" is an industrial control system range. Closer to a real intrusion, but there are no active defenders and initial access is typically already given.
What this does and does not show
- Capability on published lab evaluations is rising quickly.
- More than one developer's model has now completed a multi-step range.
- The first ICS range completion was published in May 2026.
- Peer benchmarks broaden coverage without being combined into a synthetic global score.
- It does not show real-world attack rates or success against defended networks.
- It does not show that OT is already under AI-driven attack.
- It is not a forecast. AISI notes trends are not guaranteed to continue.