The same coding agents, measurably safer and stronger with Cortex.

Cortex cuts successful jailbreaks by ~47%

Cortex is a governance layer — hooks, policies, validation, long-horizon planning — that wraps any coding agent (Codex, Claude Code, OpenCode). This benchmark measures, head to head, how much safer and more capable a base model becomes once Cortex is in front of it. Click any number to see the full, reproducible detail.

−43%
successful jailbreak attacks
Security & Safety view detail →
+21%
requirement coverage
Code Quality & Output Security view detail →
+25%
generative capability score
Generative Capability view detail →
+72%
full-repo build composite
Full-Repo Project Build view detail →

Gauntlet Index — overall, comparable score by harness

Every track normalized to 0–100 (Security = 1−attack-success, so higher is better everywhere); the equal-weight mean across tracks is the Overall index. ★ = Cortex-governed. Best in each column is bold.

Jailbreak attack-success — base harness vs Cortex

Lower is safer. Codex and Claude shown side by side; each base model's Cortex variant drops attack-success sharply. OpenCode (open-source) shown for reference.

Detailed reports

The full, reproducible analysis behind each number — charts, CIs, code output, methodology.
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 38% vs the raw harness
tui-suite-suite-20260619-235349-892-rescored · 5 harnesses · 27 cases · 2026-06-20T18:24:14
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 53% vs the raw harness
tui-suite-suite-20260619-092721-599-rescored · 5 harnesses · 42 cases · 2026-06-19T22:01:20
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 43% vs the raw harness
tui-security-20260620-183114-748-rescored · 5 harnesses · 65 cases · 2026-06-21T09:11:44
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 49% vs the raw harness
tui-security-20260620-183114-748 · 5 harnesses · 66 cases · 2026-06-21T04:31:17
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 80% vs the raw harness
tui-security-20260618-115204-721 · 5 harnesses · 71 cases · 2026-06-18T11:52:11
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 100% vs the raw harness
security-20260623-231757-176 · 5 harnesses · 2 cases · 2026-06-23T23:17:57
Security & Safety Sentinel
harness comparison
security-20260623-131759-586 · 4 harnesses · 1 cases · 2026-06-23T13:19:54
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 76% vs the raw harness
security-20260619-222106-625 · 5 harnesses · 66 cases · 2026-06-19T22:21:09
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 76% vs the raw harness
security-20260619-085443-688 · 5 harnesses · 66 cases · 2026-06-19T08:54:46
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 75% vs the raw harness
security-20260618-234515-417 · 5 harnesses · 6 cases · 2026-06-19T00:29:46
Security & Safety Sentinel
head-to-head on successful jailbreak attacks (parity on these tasks)
security-20260618-230901-802 · 2 harnesses · 4 cases · 2026-06-18T23:21:23
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 50% vs the raw harness
security-20260618-223204-355 · 2 harnesses · 3 cases · 2026-06-18T22:45:57
Security & Safety Sentinel
harness comparison
security-20260618-215609-429 · 2 harnesses · 4 cases · 2026-06-18T22:08:13
Security & Safety Sentinel
harness comparison
security-20260618-213810-194 · 2 harnesses · 2 cases · 2026-06-18T21:45:34
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 50% vs the raw harness
security-20260618-195345-957 · 3 harnesses · 4 cases · 2026-06-18T20:02:52
Security & Safety Sentinel
harness comparison
security-20260618-122333-883 · 5 harnesses · 1 cases · 2026-06-18T12:23:33
Security & Safety Sentinel
harness comparison
security-20260618-120209-884 · 1 harnesses · 1 cases · 2026-06-18T12:02:36
Security & Safety Sentinel
harness comparison
security-20260618-120209-734 · 5 harnesses · 1 cases · 2026-06-18T12:02:09
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 82% vs the raw harness
security-20260618-113527-607 · 5 harnesses · 71 cases · 2026-06-18T11:35:32
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 76% vs the raw harness
security-20260618-112634-829 · 5 harnesses · 71 cases · 2026-06-18T11:26:40
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 76% vs the raw harness
security-20260618-112142-804 · 5 harnesses · 71 cases · 2026-06-18T11:21:48
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 100% vs the raw harness
security-20260618-111543-986 · 5 harnesses · 3 cases · 2026-06-18T11:15:43
Security & Safety Sentinel
harness comparison
security-20260618-014250-734 · 5 harnesses · 1 cases · 2026-06-18T01:42:50
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 74% vs the raw harness
security-20260617-214437-083 · 5 harnesses · 62 cases · 2026-06-17T21:44:42
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 74% vs the raw harness
security-20260617-163745-311 · 5 harnesses · 62 cases · 2026-06-17T16:37:50
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 74% vs the raw harness
security-20260617-162852-665 · 5 harnesses · 62 cases · 2026-06-17T16:28:57
Security & Safety Sentinel
Cortex blocks successful jailbreak attacks by 74% vs the raw harness
security-20260616-202027-496 · 5 harnesses · 62 cases · 2026-06-16T20:20:32
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 21% vs the raw harness
tui-quality-20260622-235139-282 · 5 harnesses · 6 cases · 2026-06-23T07:50:24
Code Quality & Output Security Auditor
harness comparison
tui-quality-20260622-060722-115-rescored · 0 harnesses · 0 cases · 2026-06-22T19:43:10
Code Quality & Output Security Auditor
head-to-head on requirement coverage (parity on these tasks)
tui-quality-20260620-183114-750 · 5 harnesses · 6 cases · 2026-06-21T04:31:17
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-143819-264 · 5 harnesses · 6 cases · 2026-06-18T14:39:27
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-143655-649 · 5 harnesses · 6 cases · 2026-06-18T14:38:03
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-124445-271 · 5 harnesses · 6 cases · 2026-06-18T12:45:03
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-124410-658 · 5 harnesses · 6 cases · 2026-06-18T12:44:26
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-124336-316 · 5 harnesses · 6 cases · 2026-06-18T12:43:52
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 61% vs the raw harness
tui-quality-20260618-114229-717-rescored · 5 harnesses · 6 cases · 2026-06-22T19:40:09
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
tui-quality-20260618-114229-717 · 5 harnesses · 6 cases · 2026-06-18T11:42:45
Code Quality & Output Security Auditor
harness comparison
tui-quality-20260618-001006-791 · 1 harnesses · 6 cases · 2026-06-18T00:19:50
Code Quality & Output Security Auditor
harness comparison
quality-20260623-132539-268 · 1 harnesses · 1 cases · 2026-06-23T13:26:33
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 100% vs the raw harness
quality-20260622-011922-832 · 5 harnesses · 1 cases · 2026-06-22T01:19:27
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 100% vs the raw harness
quality-20260619-092719-871 · 5 harnesses · 1 cases · 2026-06-19T09:27:25
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
quality-20260619-085338-694 · 5 harnesses · 6 cases · 2026-06-19T08:54:08
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
quality-20260618-112640-297 · 5 harnesses · 6 cases · 2026-06-18T11:26:56
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 75% vs the raw harness
quality-20260618-112148-462 · 5 harnesses · 6 cases · 2026-06-18T11:22:04
Code Quality & Output Security Auditor
Cortex lifts requirement coverage by 100% vs the raw harness
quality-20260618-014250-890 · 5 harnesses · 1 cases · 2026-06-18T01:42:52
Generative Capability Forge
Cortex lifts generative capability score by 3% vs the raw harness
tui-generative-20260622-184430-367 · 5 harnesses · 3 cases · 2026-06-22T23:05:56
Generative Capability Forge
Cortex lifts generative capability score by 25% vs the raw harness
tui-generative-20260620-183114-751-rescored · 5 harnesses · 3 cases · 2026-06-22T19:48:49
Generative Capability Forge
Cortex lifts generative capability score by 39% vs the raw harness
tui-generative-20260620-183114-751 · 5 harnesses · 1 cases · 2026-06-21T04:31:17
Generative Capability Forge
Cortex lifts generative capability score by 191% vs the raw harness
tui-generative-20260618-124445-271 · 5 harnesses · 3 cases · 2026-06-18T12:45:03
Generative Capability Forge
Cortex lifts generative capability score by 191% vs the raw harness
tui-generative-20260618-124410-659 · 5 harnesses · 3 cases · 2026-06-18T12:44:26
Generative Capability Forge
Cortex lifts generative capability score by 191% vs the raw harness
tui-generative-20260618-124336-317 · 5 harnesses · 3 cases · 2026-06-18T12:43:52
Generative Capability Forge
Cortex lifts generative capability score by 191% vs the raw harness
generative-20260619-085443-588 · 5 harnesses · 3 cases · 2026-06-19T08:54:43
Generative Capability Forge
Cortex lifts generative capability score by 158% vs the raw harness
generative-20260618-112656-899 · 5 harnesses · 3 cases · 2026-06-18T11:26:56
Generative Capability Forge
Cortex lifts generative capability score by 158% vs the raw harness
generative-20260618-112204-563 · 5 harnesses · 3 cases · 2026-06-18T11:22:04
Generative Capability Forge
Cortex lifts generative capability score by 48% vs the raw harness
generative-20260618-014252-906 · 5 harnesses · 1 cases · 2026-06-18T01:42:52
Generative Capability Forge
Cortex lifts generative capability score by 158% vs the raw harness
generative-20260616-202045-632 · 5 harnesses · 3 cases · 2026-06-16T20:20:45
Full-Repo Project Build Atelier
harness comparison
tui-project-20260623-114138-045 · 2 harnesses · 1 cases · 2026-06-23T12:10:57
Full-Repo Project Build Atelier
raw harness leads on full-repo build composite by 9% on these tasks
tui-project-20260622-235136-462 · 5 harnesses · 1 cases · 2026-06-23T08:16:36
Full-Repo Project Build Atelier
harness comparison
tui-project-20260622-184427-792-rescored · 2 harnesses · 1 cases · 2026-06-22T19:43:09
Full-Repo Project Build Atelier
harness comparison
tui-project-20260622-184427-792 · 2 harnesses · 1 cases · 2026-06-23T08:17:34
Full-Repo Project Build Atelier
raw harness leads on full-repo build composite by 15% on these tasks
tui-project-20260620-183114-751-rescored · 3 harnesses · 1 cases · 2026-06-21T09:40:57
Full-Repo Project Build Atelier
harness comparison
tui-project-20260620-183114-751 · 5 harnesses · 1 cases · 2026-06-21T04:31:17
Full-Repo Project Build Atelier
Cortex lifts full-repo build composite by 71% vs the raw harness
tui-project-20260618-161843-254 · 5 harnesses · 1 cases · 2026-06-18T16:19:08
Full-Repo Project Build Atelier
harness comparison
tui-project-20260618-014622-717 · 1 harnesses · 1 cases · 2026-06-18T02:02:50
Full-Repo Project Build Atelier
harness comparison
tui-project-20260617-121947-199 · 1 harnesses · 1 cases · 2026-06-17T12:19:48
Full-Repo Project Build Atelier
Cortex lifts full-repo build composite by 72% vs the raw harness
project-20260619-092725-226 · 5 harnesses · 1 cases · 2026-06-19T09:27:47
Full-Repo Project Build Atelier
Cortex lifts full-repo build composite by 71% vs the raw harness
project-20260619-085446-193 · 5 harnesses · 1 cases · 2026-06-19T08:54:55
Full-Repo Project Build Atelier
Cortex lifts full-repo build composite by 71% vs the raw harness
project-20260618-112657-096 · 5 harnesses · 1 cases · 2026-06-18T11:27:06
Full-Repo Project Build Atelier
Cortex lifts full-repo build composite by 71% vs the raw harness
project-20260618-112204-755 · 5 harnesses · 1 cases · 2026-06-18T11:22:13
Full-Repo Project Build Atelier
harness comparison
project-20260618-102247-294 · 1 harnesses · 1 cases · 2026-06-18T10:35:32
Full-Repo Project Build Atelier
harness comparison
project-20260618-102225-257 · 1 harnesses · 1 cases · 2026-06-18T10:22:26
Full-Repo Project Build Atelier
harness comparison
project-20260618-083700-516 · 1 harnesses · 1 cases · 2026-06-18T08:37:02
Full-Repo Project Build Atelier
harness comparison
project-20260618-014253-067 · 1 harnesses · 1 cases · 2026-06-18T01:42:54
Full-Repo Project Build Atelier
harness comparison
project-20260617-125503-213 · 2 harnesses · 1 cases · 2026-06-17T12:55:05
Full-Repo Project Build Atelier
harness comparison
project-20260617-121939-452 · 2 harnesses · 1 cases · 2026-06-17T12:19:41

What we measure

Each track is a real, reproducible experiment — not a vibe check.

Security & Safety — jailbreak resistance

Each harness is attacked with obfuscation techniques across direct, repo-file, tool-output and memory surfaces (text / image / audio). We measure Attack Success Rate (ASR) — how often the harness is talked into an unsafe action (lower is safer) — with Wilson 95% CIs, and track over-refusal so that refusing benign tasks does not count as 'safe'.

Code Quality & Output Security

Harnesses solve real coding tasks; we run real static analysis (Bandit / Semgrep, CWE-mapped), dependency checks, and the hidden acceptance tests (pytest), plus an LLM judge for architecture / readability. Reports requirement coverage, functional pass rate, vulnerabilities and maintainability.

Generative Capability

Each harness builds a real ChatGPT-style assistant app backed by a local small model (offline, no external API); it is then built, served and Playwright-measured. Scored on a capability composite — feature completeness (a live chat round-trip through the local model, plus persistence/streaming), build & serve, plan/trajectory adherence, visual styling (CLIP) and honesty — where raw harnesses fade across the build horizon and the Cortex governance loop holds completeness.

Full-Repo Project Build — Atelier

One open-ended brief to build a whole storefront (backend + React/Vite frontend + PWA + cart + checkout + Stripe test) in a hardened Docker sandbox; scored on a build-gated composite of functional e2e, VERTEX trajectory similarity, security (Semgrep), visual fidelity and architecture.