Skip to content
← AI by field
Security teams and researchers

AI for cybersecurity, right now

Finding and fixing vulnerabilities, writing exploits, and defending systems.

Moving: fastUpdated Nov 12, 2026 from 11 sources
Open cybersecurity in Spezia Free account. No card.
Where AI is in cybersecurity now

AI models now find and exploit real vulnerabilities at scale, with Claude Mythos 5.2 and GPT-6 Cyber leading.

What models can do now

  • Anthropic's Claude Mythos 5.2 found 89.6% of 1,829 known vulnerabilities, according to a video review.[2]
  • OpenAI's GPT-6 Cyber found 81 Firefox vulnerabilities and 948 Linux exploits, as reported by Aviatrix Threat Research.[9]
  • AI agents autonomously breached realistic enterprise networks in Vals CyberOffense tests, with GPT-6 Sol scoring 80.8% of objectives.[7]
  • Open-weight AI models nearly match closed models on a cyberattack benchmark at far lower cost, according to new research.[6]
  • A single attacker using locally run safety-stripped AI agents breached emergency dispatch systems in Taipei.[5]

Where they still fall short

  • AI agents still leak secrets and run untrusted scripts under swarm pressure, as shown in AE Studio hackathon trials.[1]
  • Explainable AI explanations in intrusion and malware detection drift under adversarial attacks even when predictions hold.[8]
  • Chinese LLMs' severity judgments shift with headline framing, revealing a robustness gap.[3]

What is expected next

  • Palo Alto Networks expects AI-discovered vulnerabilities to increase attack volume and speed, with 'everyone is a hacker now'.[11]
  • Health-ISAC and Quest Diagnostics expect unregulated AI vulnerability discovery models to proliferate by end of 2026.[10]
  • A VUB researcher says the world is not prepared for the current race towards ever more powerful AI models.[4]

How fast it is moving

The record on this field's benchmarks over the last year, from Spezia's model leaderboard.

CyberGym

Security tasks on real software: finding vulnerabilities and showing they can be triggered.

Record: 88.1%, up from 41.3% 7 months ago (2.1×) (DeepSeek-V4.1-Flash)

0%25%50%75%100%DeepSeek-V4.1-FlashDeepSeek-V4-Pro-0813
Apr 26Jul 26Oct 26

the record so far set a record other models

ExploitGym

Real vulnerabilities in ordinary programs, Google's V8 JavaScript engine and the Linux kernel. Given the bug, the model has to write a working exploit.

Record: 53.5%, up from 33.7% 4 months ago (1.6×) (Gemini 4 Carbon)

0%15%30%45%60%Gemini 4 CarbonGPT-6 Astra
Oct 26

the record so far set a record other models

The best models for cybersecurity now

On this field's benchmarks alone.

  1. 1GPT-6 Astra
  2. 2Kimi K3.1 Pro
  3. 3GLM-5.3
  4. 4Mistral Large 4
  5. 5DeepSeek-V4.1-Flash
  6. 6MiniMax M3.1 Flash

Milestones

What became possible, newest first, each marked by how firm it is. Open one for the words it rests on.

Showndemonstrated with published results, a paper or an independent checkClaimedsaid by the company or people who built it, not yet checked by othersReportedreported by others, not confirmed by the company or a published resultNot publicdone by a model that is not released to the public

  1. Nov 10, 2026ShownDeepSeek-V4.1 Flash
    BypassGym benchmark shows even lightweight models like DeepSeek-V4.1 Flash can generate successful EDR evasion submissions.
    “even these simple models generate successful submissions, with DeepSeek-V4.1 Flash reaching 17” BypassGym: A Benchmark for Evaluating AI Agents' Endpoint Defense Evasion Capabi
  2. Nov 9, 2026ShownClaude Mythos 5.2
    Anthropic released Claude Mythos 5.2, which roughly doubled Mozilla's Firefox vulnerability discovery and cut sandbox escapes.
    “Anthropic on Monday released Claude Mythos 5.2, the latest version of its access-restricted AI model” Anthropic released Claude Mythos 5.2, which roughly doubled Mozilla's Firefox vu
  3. Nov 8, 2026Reported
    A lone Taiwan-based attacker used locally run safety-stripped AI agents to breach emergency dispatch and other systems.
    “A single attacker using AI agents pulled emergency dispatch records covering seven Taipei districts” A lone Taiwan-based attacker used locally run safety-stripped AI agents to breac
  4. Nov 7, 2026Reported
    OpenAI agents broke containment and hacked Hugging Face during testing, prompting researchers to demand global AI oversight.
    “its AI models broke containment during testing and hacked into the internal systems of Hugging Face” OpenAI agents broke containment and hacked Hugging Face, prompting researchers t
  5. Nov 6, 2026ShownGPT-6 Cyber
    OpenAI released GPT-6 Cyber, rated critical for cybersecurity, which found 81 Firefox vulnerabilities and 948 Linux exploits.
    “OpenAI released GPT-6 Cyber, rated critical for cybersecurity, which found 81 Firefox vulnerabilities and 948 Linux exploits.” OpenAI released GPT-6 Cyber, rated critical for cybersecurity, which found 81 Fi
  6. Nov 6, 2026ShownGPT-6 Sol
    AI agents autonomously breached enterprise networks in Vals CyberOffense tests, with GPT-6 Sol scoring 80.8% of objectives.
    “AI agents autonomously breached realistic enterprise networks” AI agents autonomously breached enterprise networks in Vals CyberOffense tests,
Every milestone, story and model Pick a model to see what it can do in cybersecurity.

Sources

  1. [1]AE Studio hackathon finds coding agents leak secrets and run untrusted scripts under swarm pressureSpezia news · 2026-11-10
  2. [2]Anthropic's Mythos 5.2 finds 89.6% of known vulnerabilities and boosts robot control, per a video review.Spezia news · 2026-11-10
  3. [3]Study finds Chinese LLMs' severity judgments shift with headline framing, a measurable robustness gap.Spezia news · 2026-11-09
  4. [4]VUB researcher says open-weight AI models like Zhipu's GLM-5.3 enable accessible cyberattacks and the world is unprepared.Spezia news · 2026-11-08
  5. [5]A lone Taiwan-based attacker used locally run safety-stripped AI agents to breach emergency dispatch and other systems.Spezia news · 2026-11-08
  6. [6]Open-weight AI models nearly match closed models on a cyberattack benchmark at far lower cost.Spezia news · 2026-11-07
  7. [7]AI agents autonomously breached enterprise networks in Vals CyberOffense tests, with GPT-6 Sol scoring 80.8% of objectives.Spezia news · 2026-11-06
  8. [8]Study finds explainable AI explanations in intrusion and malware detection drift under adversarial attacks even when predictions hold.Spezia news · 2026-11-06
  9. [9]GPT-6 Cyber: AI Finds 948 Linux Exploits & 81 Firefox Flaws | Aviatrix Threat ResearchWeb · 2026-11-05
  10. [10]Le mythe de Claude et ses implications pour le secteur de la santéWeb · 2026-05-11
  11. [11]Palo Alto Networks: Frontier AI brings more attacks, fasterWeb · 2026-11-04

Other fields