CyberGym's own paper put the best agent at 11.9% (early version) to 22.0% (final version). Vendors now claim 93–97% on the same suite: TrendAI at 97% and a self-reported 95.1% for Xiaomi's MiMo-V2.6-Flash.
Best AI for Cybersecurity 2026: Exploit vs Fix Scores
In CyberGym's own paper the best AI agent reproduced at most 22% of real vulnerabilities, and vendors now claim 95–97% on the same 1,507 bugs.
- No. 1: GPT-6 Sol (OpenAI)
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
GPT-6 Sol (OpenAI)
Second on CyberBench v1.1 at 77.98%, and about 69% on the Artificial Analysis Cyber Index in its gated Daybreak Blue form. No fix-benchmark score was found.
Included it: Claude, Meta, Grok, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, MiMo, Kimi, Wenxin.In their lists: Claude: 6 · Meta: 2 · Grok: see list · DeepSeek: 5
What each AI noted 4
- Claudeplace 6
Second on CyberBench v1.1, with no SWE-bench figure in my sources.
- Metaplace 2
78.0% on CyberBench and the gated Daybreak Blue configuration is the benchmark for what attackers would want — if they could get it.
- Grok
Named differently in its own list: open the full list.
see list - DeepSeekplace 5
Second on CyberBench (77.98%) with strong reasoning, but its exploitation-to-fix balance is less proven than Mythos or Gemini Cyber.
Source ↗
4 AIs put it here - 2
Muse Wonder (Meta)
Leads Vals CyberBench v1.1 at 80.36% and scores 96.7% on SWE-bench Pro V2 Full, so it is strong at both breaking and fixing. A 'Wonder Full' listing shows 73.6%.
Included it: Claude, Meta, Grok, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, MiMo, Kimi, Wenxin.In their lists: Claude: 7 · Meta: 1 · Grok: 7 · DeepSeek: 4
What each AI noted 4
- Claudeplace 7
Top of CyberBench v1.1 at 80.36%, but I found no fix scores for it.
- Metaplace 1
The only model that tops both Vals CyberBench at 80.4% and Scale's SWE-bench Pro V2 Full at 96.7% — exploit reproduction and fixing in the same system.
Source ↗ - Grokplace 7
Leads one CyberBench snapshot at 80.36% and strong on combined find-and-patch tasks
- DeepSeekplace 4
Leads CyberBench at 80.36% and scores 88.4% on CyberGym, but its SWE-bench fix profile is less established.
Source ↗
4 AIs put it here - 3
Gemini 4 Argon (Google)
Scores 77.86% on CyberBench, 0.12 points behind GPT-6 Sol, plus 65–68% on CWE-bench patching. It is one of the few models scored on both sides.
Included it: Claude, Meta, Grok, DeepSeek. Did not include: ChatGPT, Nemotron, Gemini, MiMo, Kimi, Wenxin.In their lists: Claude: 5 · Meta: 3 · Grok: 3 · DeepSeek: 10
What each AI noted 4
- Claudeplace 5
Only 0.12 points behind GPT-6 Sol on CyberBench, from a company that also builds the defender tools Big Sleep and CodeMender.
- Metaplace 3
77.9% on CyberBench and 68% on CWE-bench v1 — the strongest defence-leaning Google model in this set.
- Grokplace 3
Strong across CyberBench (~78%) and CWE-bench patching; Carbon leads some exploit-oriented boards
- DeepSeekplace 10
Third on CyberBench at 77.86%, with high reasoning effort, but its CyberGym and SWE-bench balance is less documented than the leaders.
Source ↗
4 AIs put it here - 4
DeepSeek V4 (Pro and V4.1 Flash)
Open-weight: V4 Pro fixes 96.40% on SWE-bench Verified, and V4.1 Flash scores 73.69% on CyberBench at under $3 a run, against about $22 for GPT-6 Sol.
Included it: Claude, Grok, Kimi, DeepSeek. Did not include: ChatGPT, Meta, Nemotron, Gemini, MiMo, Wenxin.In their lists: Claude: 3 · Grok: 4 · Kimi: 9 · DeepSeek: 7
What each AI noted 4
- Claudeplace 3
Second on SWE-bench Verified at 96.40%, an open-weight-class model that attackers can run themselves.
- Grokplace 4
Open-weight model closes within ~4-8 points of gated frontier systems on CyberBench and multi-step offense at a fraction of the cost
- Kimiplace 9
73.69% on Vals.ai CyberBench, edging GLM-5.3 in less than half the runtime — price doesn't dictate cyber capability.
Source ↗ - DeepSeekplace 7
83.3% on CyberGym and competitive SWE-bench performance at lower cost, making it a value option for exploit-heavy work.
Source ↗
4 AIs put it here - 5
Claude Opus 5 (Anthropic)
The best measured fixer on two publishers' SWE-bench Verified tables (97.00% at Vals, 96% at BenchLM). Vals calls that test saturated, and no exploit score exists for this model.
Included it: Claude, MiMo, Kimi. Did not include: ChatGPT, Meta, Grok, Nemotron, Gemini, DeepSeek, Wenxin.In their lists: Claude: 1 · MiMo: 1 · Kimi: 3
What each AI noted 3
- Claudeplace 1
First on SWE-bench Verified at 97.00%, so it fixes more than it breaks, and Google's tracker did not name Claude among attacker tools.
- MiMoplace 1
96% on SWE-bench Verified — the best fixing score measured, which is exactly the skill defenders need and attackers use.
Source ↗ - Kimiplace 3
96% on SWE-bench Verified — the highest rate of real bug fixes that pass without breaking other tests.
Source ↗
3 AIs put it here - 6
Claude Fable 5.2 (Anthropic)
Top patcher: 75% on CWE-bench v2, run by the benchmark's maintainers, 7 points ahead of Gemini 4 Hydrogen. It also scores 93% on SWE-bench Pro, and no exploit score is published.
Included it: Meta, Grok, Kimi. Did not include: Claude, ChatGPT, Nemotron, Gemini, MiMo, DeepSeek, Wenxin.In their lists: Meta: 5 · Grok: 1 · Kimi: 1
What each AI noted 3
- Metaplace 5
Leads CWE-bench v2 at 75% and SWE-bench Pro at 93% in one November snapshot — the patching specialist of the Claude line.
Source ↗ - Grokplace 1
Leads current CWE-bench patching at 75% while Mythos variants set early high-water marks for end-to-end exploit construction
- Kimiplace 1
The best published fixer: 75% on CWE-bench v2, the top vulnerability-patching score — and it was run by the benchmark's own maintainers, not the lab.
Source ↗
3 AIs put it here - 7
Claude Mythos 5 (Anthropic)
Fixes 95.5% on SWE-bench Verified but reaches only 38.1% on ExploitGym, against GPT-6 Cyber's 48.5%. The profile leans to defence, though each lab used its own harness.
Included it: Claude, MiMo, Kimi. Did not include: ChatGPT, Meta, Grok, Nemotron, Gemini, DeepSeek, Wenxin.In their lists: Claude: 4 · MiMo: 2 · Kimi: 6
What each AI noted 3
- Claudeplace 4
Second on BenchLM's fix table at 95.5%, but I found no Mythos-specific exploit numbers.
Source ↗ - MiMoplace 2
95.5% — a half point behind its sibling, and together the top three of one family sweep the entire fix leaderboard.
Source ↗ - Kimiplace 6
Defense-leaning profile: 95.5% fixing on SWE-bench Verified versus 38.1% exploit reproduction on ExploitGym.
Source ↗
3 AIs put it here - 8
GPT-6 Cyber (OpenAI)
An exploit specialist: 48.5% on ExploitGym and 100% on ExploitBench, both reported by OpenAI. It is available only behind identity checks, and no fix score is published.
Included it: Meta, Grok, Kimi. Did not include: Claude, ChatGPT, Nemotron, Gemini, MiMo, DeepSeek, Wenxin.In their lists: Meta: 8 · Grok: 2 · Kimi: 4
What each AI noted 3
- Metaplace 8
48.5% on OpenAI's own ExploitGym — a specialist that is deliberately gated because its exploit skill is the point.
Source ↗ - Grokplace 2
48.5% on ExploitGym (51% relative gain over prior Cyber) and strong multi-step offensive completion; restricted access limits open use
- Kimiplace 4
100% on ExploitBench and the best black-box web pentest score (40.3%) — and you can only use it behind identity verification, SSO and IP allowlists.
Source ↗
3 AIs put it here - 9
Claude Opus 5.5 (Anthropic)
Tops Vals CyberOffense network attacks at 77.1% and fixes 89.9% on SWE-bench Pro. Its CyberBench score reads 55.36% or 74.6%, depending on the Vals table.
Included it: Meta, Grok, DeepSeek. Did not include: Claude, ChatGPT, Nemotron, Gemini, MiMo, Kimi, Wenxin.In their lists: Meta: 4 · Grok: 8 · DeepSeek: 9
What each AI noted 3
- Metaplace 4
The best multi-stage attacker in Vals CyberOffense at 77.1% — and also a top fixer at 89.9% on SWE-bench Pro.
- Grokplace 8
High SWE-bench fixing rates and solid CWE-bench; heavier refusal rates on pure offensive tasks without trusted-access routing
- DeepSeekplace 9
73.1% on CyberGym and 64.3% on SWE-bench Pro — stronger on fixes than pure exploit reproduction, but a step behind Mythos.
Source ↗
3 AIs put it here - 10
Kimi K3.1 Pro (Moonshot AI)
Highest Artificial Analysis Cyber Index score on a standard public API (64.1%), and 94.7% on CyberGym-E2E-AA, above gated GPT-6 Sol Daybreak Blue's 93%.
Included it: Meta, Grok. Did not include: Claude, ChatGPT, Nemotron, Gemini, MiMo, Kimi, DeepSeek, Wenxin.In their lists: Meta: 6 · Grok: 6
What each AI noted 2
- Metaplace 6
The highest Cyber Index score available through a standard public API at 64.1% — and 94.7% on CyberGym-E2E-AA.
Source ↗ - Grokplace 6
Top independent CyberGym end-to-end scores and competitive multi-step offensive results
2 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
The harness moves scores as much as the model does: in one case Semgrep saw a model go from 25% to 73% with a better harness.
Test versions are not comparable: GPT-5.6 Sol dropped from 88.14% on CyberBench v1 to 76.31% on v1.1. Vals also counts refusals as failures.
Fix scores are inflated too: OpenAI stopped reporting SWE-bench Verified after an audit found material test flaws in 59.4% of the hard tasks it checked.
Attackers in the field run open-source agent frameworks (ARTEX, CAIRN, PentAGI, CyberStrikeAI, exploit_agent), not the models topping the leaderboards. The FBI says ARTEX reached privileged access in under three hours.
03Questions people ask
What is the best AI for cybersecurity?
No single model wins. Muse Wonder leads Vals CyberBench v1.1 at 80.36%, while Claude Fable 5.2 leads CWE-bench v2 patching at 75%.
What is the best AI model for cybersecurity?
It depends on the job. For fixing real bugs, Claude Opus 5 leads SWE-bench Verified at 97.00%; for simulated network attacks, Claude Opus 5.5 leads Vals CyberOffense at 77.1%.
What is the best AI cybersecurity stock?
The benchmarks rank models, not stocks. Wedbush named CrowdStrike ($250 target) and Palo Alto Networks ($400 target) as its top cybersecurity picks.
Source ↗What is the best AI cybersecurity certification?
CompTIA SecAI+ (CY0-001), launched 17 Feb 2026, is the AI-specific starting point: up to 60 questions in 60 minutes, aimed at people with a few years of security work.
Source ↗04How we counted
In plain words, no formulas.
1The question
Top 10 AI Models for Cybersecurity: Exploit Skill vs. Fixing Skill
2Why this topic
Readers see attack-side scores, defense-side scores and what was observed in real attacks together. They learn which models help defenders most, and why the same test gives very different answers depending on who runs it.
3Where it applies
Global; frontier AI models and agent setups tested on real software flaws
4What we looked at
- Success rate at reproducing real vulnerabilities in the lab (CyberGym, Vals.ai CyberBench)
- Success rate at producing working, regression-free fixes for real code bugs (SWE-bench family)
- Whether a human is needed to finish the job
- How consistent a model's scores are across test versions and publishers
- Vals.ai CyberBench v1.1 — Proof-of-concept reproduction (60 tasks) and regression-safe patches (56 tasks) on real OSS-Fuzz bugs; not comparable with v1
- CWE-bench v2 (via BenchLM) — Patching of weakness-typed vulnerabilities, run by the benchmark's maintainers
- SWE-bench Verified (Vals, BenchLM) — Share of real open-source bugs fixed without breaking other tests; Vals marks it saturated
- SWE-bench Pro (Scale AI SEAL) — Harder real-repository fixing tasks; V2 uses 642 tasks
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Nemotron, Gemini, MiMo, Kimi, DeepSeek, Wenxin.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Claude Opus 5
First on SWE-bench Verified at 97.00%, so it fixes more than it breaks, and Google's tracker did not name Claude among attacker tools.
- 2GPT-5.6 Sol
Fell from 88.14% on CyberBench v1 to 76.31% on v1.1, which shows how much the test version matters.
- 3DeepSeek V4 Pro
Second on SWE-bench Verified at 96.40%, an open-weight-class model that attackers can run themselves.
- 4Claude Mythos 5
Second on BenchLM's fix table at 95.5%, but I found no Mythos-specific exploit numbers.
Source ↗ - 5Gemini 4 Argon
Only 0.12 points behind GPT-6 Sol on CyberBench, from a company that also builds the defender tools Big Sleep and CodeMender.
- 6GPT-6 Sol
Second on CyberBench v1.1, with no SWE-bench figure in my sources.
- 7Wonder (agent setup)
Top of CyberBench v1.1 at 80.36%, but I found no fix scores for it.
- 8Claude Fable 5
95% on BenchLM's fix table, close to the Opus and Mythos models.
Source ↗ - 9Sangfor AI Agent on DeepSeek-V4-Flash
Reported 93.2% on CyberGym from one report, which sits oddly beside a verified 63.24% in a September paper.
Source ↗ - 10Microsoft MDASH agent
Reported 88.45% on CyberGym in May 2026.
Source ↗
What Claude noticed
CyberGym's early scores were low: 17.9% for the best single-trial result in the paper. GPT-5 rose from 7.7% to 22.0% with more reasoning effort.
The agent harness can move results as much as the model does: Semgrep saw one model go from 25% to 73% with a better harness.
Berkeley RDI reported Claude Sonnet 4.5 at 28.9% on CyberGym in a single run, rising to 66.7% with 30 trials.
GTIG says attackers are using AI mostly for n-day work, such as analyzing patches and proof-of-concept code. It reports that China-linked UNC2814 and North Korea's APT45 used AI for vulnerability research.
🇺🇸 Meta10 places · 3 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Muse Wonder (Meta)
The only model that tops both Vals CyberBench at 80.4% and Scale's SWE-bench Pro V2 Full at 96.7% — exploit reproduction and fixing in the same system.
Source ↗ - 2GPT-6 Sol (OpenAI)
78.0% on CyberBench and the gated Daybreak Blue configuration is the benchmark for what attackers would want — if they could get it.
- 3Gemini 4 Argon (Google)
77.9% on CyberBench and 68% on CWE-bench v1 — the strongest defence-leaning Google model in this set.
- 4Claude Opus 5.5 (Anthropic)
The best multi-stage attacker in Vals CyberOffense at 77.1% — and also a top fixer at 89.9% on SWE-bench Pro.
- 5Claude Fable 5.2 (Anthropic)
Leads CWE-bench v2 at 75% and SWE-bench Pro at 93% in one November snapshot — the patching specialist of the Claude line.
Source ↗ - 6Kimi K3.1 Pro (Moonshot AI)
The highest Cyber Index score available through a standard public API at 64.1% — and 94.7% on CyberGym-E2E-AA.
Source ↗ - 7TrendAI AESIR (Trend Micro agent setup)
97% on CyberGym — not a raw model, but the highest-scoring exploit-remediation engine in the set.
Source ↗ - 8GPT-6 Cyber (OpenAI)
48.5% on OpenAI's own ExploitGym — a specialist that is deliberately gated because its exploit skill is the point.
Source ↗ - 9GLM-5.3 (Z.ai)
84.5% on CyberGym in Z.ai's own run — a reminder that open-weight models now sit near the frontier on exploit tasks.
Source ↗ - 10Grok 4.7 (xAI)
56.4% on the Artificial Analysis Cyber Index — strong, but its CyberGym-E2E-AA score trails Kimi by more than 20 points.
Source ↗
What Meta noticed
The lab leader is not the attacker tool GTIG actually saw: GTIG names ARTEX, CAIRN, CyberStrikeAI, PentAGI and exploit_agent in the wild, not the benchmark-topping models by name.
Kimi K3.1 Pro shows the open-weight twist: 95% on CyberGym-E2E-AA beats the gated GPT-6 Sol Daybreak Blue configuration at 93% on that sub-test.
TrendAI's 97% on CyberGym is an agent system, not a base model — proof that harness, tools and remediation workflow can beat a bigger raw model.
🇺🇸 Grok10 places
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Claude Fable 5.2 / Mythos lineage (Anthropic)
Leads current CWE-bench patching at 75% while Mythos variants set early high-water marks for end-to-end exploit construction
- 2GPT-6 Cyber / Sol (OpenAI Daybreak)
48.5% on ExploitGym (51% relative gain over prior Cyber) and strong multi-step offensive completion; restricted access limits open use
- 3Gemini 4 Argon / Carbon / Hydrogen (Google)
Strong across CyberBench (~78%) and CWE-bench patching; Carbon leads some exploit-oriented boards
- 4DeepSeek V4.1 Flash 1017
Open-weight model closes within ~4-8 points of gated frontier systems on CyberBench and multi-step offense at a fraction of the cost
- 5MiMo-V2.6-Flash / Pro (Xiaomi)
Highest reported CyberGym reproduction rates (94-95%) among published leaderboard entries
- 6Kimi K3 / K3.1 Pro (Moonshot)
Top independent CyberGym end-to-end scores and competitive multi-step offensive results
- 7Muse Wonder (Meta)
Leads one CyberBench snapshot at 80.36% and strong on combined find-and-patch tasks
- 8Claude Opus 5.5 / Sonnet 5.5 lineage
High SWE-bench fixing rates and solid CWE-bench; heavier refusal rates on pure offensive tasks without trusted-access routing
- 9Grok 4.7 (xAI)
Competitive mid-pack CyberGym and CyberBench scores with practical agentic tooling for both offense and defense
- 10GLM-5.3 / Mistral Large 4
Open or lightly restricted models posting 70-84% CyberGym-style reproduction and solid patch rates at lower cost
🇺🇸 Gemini10 places · 1 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Claude 3.5 Sonnet (Anthropic)
Leads all frontier models in patch generation with a 51.4% SWE-bench Verified score while maintaining a defensive surplus over exploit reproduction.
Source ↗ - 2OpenAI o1 (OpenAI)
Highest raw reasoning score for deconstructing obfuscated vulnerability patches (42.6% CyberGym), but throttled by aggressive post-training safety guardrails.
Source ↗ - 3GPT-4o (OpenAI)
The workhorse model most frequently integrated into both SOC defensive triages and adversary automated patch-diffing pipelines.
Source ↗ - 4Gemini 1.5 Pro (Google DeepMind)
Massive 2M context window enables whole-codebase AST parsing, excelling at architectural vulnerability discovery and multi-repo patch synthesis.
Source ↗ - 5DeepSeek Coder V2 (DeepSeek)
The premier open-weights model for offensive diffing, scoring 31.5% on SWE-bench and heavily favored by threat actors avoiding hosted API audit logs.
Source ↗ - 6Llama 3.1 405B (Meta)
Uncensored local enterprise deployments deliver 29.2% SWE-bench fixing while powering automated fuzzing harnesses in closed labs.
- 7Qwen 2.5 72B Instruct (Alibaba Cloud)
Achieves high code generation efficiency and 32.1% SWE-bench resolution at a fraction of frontier compute overhead.
Source ↗ - 8Claude 3 Opus (Anthropic)
Exhibits superior static audit reasoning and nuanced logic error detection, though slower execution speeds limit automated loop agility.
Source ↗ - 9Mistral Large 2 (Mistral AI)
Demonstrates balanced 26.8% SWE-bench capability with strong native C/C++ memory-safety comprehension.
Source ↗ - 10GPT-4o mini (OpenAI)
High-throughput, low-latency API architecture makes it the favorite utility model for mass commit triage and initial diff screening.
Source ↗
What Gemini noticed
Google Threat Intelligence documented the first fully autonomous weaponization pipeline (exploit_agent), where an AI agent ingested patch advisories and successfully deployed exploits with zero human intervention.
🇨🇳 MiMo10 places · 4 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1Claude Opus 5
96% on SWE-bench Verified — the best fixing score measured, which is exactly the skill defenders need and attackers use.
Source ↗ - 2Claude Mythos 5
95.5% — a half point behind its sibling, and together the top three of one family sweep the entire fix leaderboard.
Source ↗ - 3Claude Fable 5
95% — third place inside one vendor's lineup means the defensible edge is now a family trait, not a model.
Source ↗ - 4Gemini 3 Flash
Top-three on the Feb 2026 fix board while DeepMind's own vuln-hunting agent finds unknown bugs — the only family with a foot on both benches.
Source ↗ - 5MiniMax M2.5
A 229B Chinese model released mid-February landed top-three on SWE-bench within days — proof that the fix-skill gap is closing fastest outside the US.
Source ↗ - 6GPT (OpenAI frontier)
The attacker's favorite building block by field reports: exploit pipelines are documented on frontier coding models even where the fix leaderboard ranks it lower.
Source ↗ - 7Kimi K2.5
On the same Feb 2026 fix board as the US leaders — open-weight models now pass the bar that matters for regression-free fixes.
Source ↗ - 8DeepSeek
The budget frontier: named on the Feb 2026 fix leaderboard at a fraction of frontier pricing, which is what makes it the defender's value pick.
Source ↗ - 9GLM-5
Open-weight, on the board, and the kind of model an SMB can actually run for patch triage — availability is itself a security measure.
Source ↗ - 10ARTEX (offensive agent setup)
Not a model but the wake-up call: an open-source agent framework that reached privileged access in under three hours — lab skill has already gone operational.
Source ↗
What MiMo noticed
No single model wins cybersecurity — BenchLM's November 2026 task map shows different leaders for patching, exploit development and pentesting, so 'best AI for security' is a task question, not a brand question.
The fix bench is monopolized while the exploit bench is democratized: one vendor holds the top three SWE-bench scores (96/95.5/95%), yet the field shows attackers composing cheap open frameworks (ARTEX, PentAGI) into working intrusion pipelines.
CyberGym's scale is the reality check: 1,507 real vulnerabilities across 188 projects — reproduction of real flaws is far below lab coding scores, and the gap is where attackers live.
The FBI's three-hour privileged-access clock for ARTEX is the first hard number connecting 'model capability' to 'time inside your network' — defenders now measure in minutes, not patches.
🇨🇳 Kimi10 places · 3 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Claude Fable 5.2 (Anthropic)
The best published fixer: 75% on CWE-bench v2, the top vulnerability-patching score — and it was run by the benchmark's own maintainers, not the lab.
Source ↗ - 2Gemini 4 Carbon (Google)
The best published breaker: 53.5% on ExploitGym, the top exploit-development score — while also posting 0% success for prompt-injection attacks against it.
Source ↗ - 3Claude Opus 5 (Anthropic)
96% on SWE-bench Verified — the highest rate of real bug fixes that pass without breaking other tests.
Source ↗ - 4GPT-6 Cyber (OpenAI, Daybreak Red)
100% on ExploitBench and the best black-box web pentest score (40.3%) — and you can only use it behind identity verification, SSO and IP allowlists.
Source ↗ - 5Gemini 4 Hydrogen (Google)
Second-best patcher at 68.1% on CWE-bench v2 — 7 points behind Claude Fable 5.2 on the same independently run test.
Source ↗ - 6Claude Mythos 5 (Anthropic)
Defense-leaning profile: 95.5% fixing on SWE-bench Verified versus 38.1% exploit reproduction on ExploitGym.
Source ↗ - 7GPT-6 Astra (OpenAI)
The most-tested model here: public results on 11 cyber benchmarks, including a tied 100% on ExploitBench.
Source ↗ - 8MiMo-V2.6-Flash (Xiaomi, open-weight)
Claims the highest CyberGym score ever at 95.1% — but the number comes from Xiaomi itself, and CyberGym's original best was 11.9%.
Source ↗ - 9DeepSeek V4.1 Flash (open-weight)
73.69% on Vals.ai CyberBench, edging GLM-5.3 in less than half the runtime — price doesn't dictate cyber capability.
Source ↗ - 10GLM-5.3 (Zhipu)
72.08% on CyberBench and the lead on long-horizon CLI work in Terminal-Bench 4.0 — strong where tasks run long.
Source ↗
What Kimi noticed
CyberGym's agents didn't just reproduce known bugs — they discovered 34 zero-days that had survived undiscovered for an average of 969 days, plus 18 historically incomplete patches. The defense value of 'exploit skill' is real.
Score inflation is the story: CyberGym's own 2025 paper measured the best agent+model at 11.9% reproduction, while Xiaomi now reports 95.1% for MiMo-V2.6-Flash — a gap that reflects harnesses, trials and self-reporting as much as capability.
What attackers actually run isn't a frontier lab model — Google and Mandiant observed intrusions built on open-source agentic frameworks (ARTEX, PentAGI, CAIRN), with one ARTEX burst logging 27 events in 23 seconds. Lab leaderboards and field tooling barely overlap.
🇨🇳 DeepSeek10 places · 3 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Claude Mythos Preview (Anthropic, 2026)
Top combined exploit-and-fix profile: 83% on CyberGym and 77.8% on SWE-bench Pro, with autonomous end-to-end intrusion capability.
Source ↗ - 2GPT-5.5-Cyber (OpenAI, 2026)
Highest CyberGym score among general-purpose frontier models (85.6%), but its SWE-bench Pro fix rate (58.6%) trails Mythos.
Source ↗ - 3Gemini 3.8 Flash Cyber (Google, 2026)
86.3% on CyberGym and 73.7% on DeepSWE v1.1 — strong on both exploitation and long-horizon software fixes.
- 4Muse Wonder Full (Meta, 2026)
Leads CyberBench at 80.36% and scores 88.4% on CyberGym, but its SWE-bench fix profile is less established.
Source ↗ - 5GPT-6 Sol (OpenAI, 2026)
Second on CyberBench (77.98%) with strong reasoning, but its exploitation-to-fix balance is less proven than Mythos or Gemini Cyber.
Source ↗ - 6XekRung-1.5-27B-Preview (Alibaba Security, 2026)
Tops CyberGym at 88.9%, the highest published reproduction score, but it is a security-specialized model with limited public SWE-bench data.
Source ↗ - 7DeepSeek V4 Pro (DeepSeek, 2026)
83.3% on CyberGym and competitive SWE-bench performance at lower cost, making it a value option for exploit-heavy work.
Source ↗ - 8GLM-5.3 (Zhipu AI, 2026)
84.5% on CyberGym and 58.4% on SWE-bench Pro — a balanced open-weight option that fixes as well as it exploits.
Source ↗ - 9Claude Opus 5.5 / Opus 4.7 (Anthropic, 2026)
73.1% on CyberGym and 64.3% on SWE-bench Pro — stronger on fixes than pure exploit reproduction, but a step behind Mythos.
Source ↗ - 10Gemini 4 Argon (Google, 2026)
Third on CyberBench at 77.86%, with high reasoning effort, but its CyberGym and SWE-bench balance is less documented than the leaders.
Source ↗
What DeepSeek noticed
CyberGym scores near 95% are raising saturation concerns and questions about whether benchmark tasks leak into training data.
OpenAI stopped reporting SWE-bench Verified in February 2026 after an audit found 59.4% of audited hard tasks had material test flaws; fix scores should be read with that contamination caveat.
The offensive-defensive asymmetry is structural: attackers need one working exploit, while defenders must find and patch every flaw, so a model that is slightly better at exploitation can shift the balance.
🇨🇳 Wenxin10 places · 1 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1NinjaOne 2026 AI Patch Automation Model
89% SWE-bench fix success rate, 47% exploit reproduction rate, defender utility 2x higher than attacker utility
Source ↗ - 2Google Threat Intelligence Mandiant LLM Security Model 2026
78% fix success rate, 52% exploit reproduction rate, observed actively used by both attack and defense teams
Source ↗ - 3ManageEngine Patch Manager Plus AI 2026
72% SWE-bench fix success rate, 32% exploit reproduction rate, fully integrated with end-to-end patching workflows
- 4Automox Cloud AI Patch Model 2026
68% fix success rate, 29% exploit reproduction rate, built for fully distributed remote workforce multi-platform patching
- 5Microsoft Intune AI Update Orchestration 2026
65% fix success rate, 27% exploit reproduction rate, purpose-built for policy-driven staged Windows update rollouts
Source ↗ - 6Patch My PC AI Vulnerability Scanner 2026
61% fix success rate, 35% exploit reproduction rate, focused on third party application patch coverage
- 7Pulseway AI Endpoint Security Model 2026
57% fix success rate, 31% exploit reproduction rate, optimized for remote MSP vulnerability response workflows
- 8Atera AI Patch Automation 2026
52% fix success rate, 28% exploit reproduction rate, built for full MSP remote monitoring and patch management
- 9Datto AI Vulnerability Response 2026
48% fix success rate, 25% exploit reproduction rate, targeted at small to mid-size MSP customer environments
- 10N-able N-central AI Security Model 2026
44% fix success rate, 22% exploit reproduction rate, integrated with full N-central remote monitoring stack
Source ↗
What Wenxin noticed
Attackers using AI agent chains to generate exploits from patches can launch over 500,000 LLM API requests in under 8 hours to support stolen token forwarding services