Why Knolt

Frontier AI, beaten on security incidents.

On Security-Bench 1.3, Knolt-Talos 1.1 made the right call 83% of the time and caught 75% of incidents. Claude Opus 5.5: 80% and 69%. It decides in 0.24 s, on your own GPU.

Security-Bench 1.3 · November 2026
  • 0%correct incident callsBest of 4Next bestClaude Opus 5.5 · 80%
  • 0%of incidents caughtBest of 4Next bestClaude Opus 5.5 · 69%
  • 0×faster than GPT-6.1 SolFastest of 4Per decision0.24 s against 2.81 s

Decides in a blink.

Each lane runs for its model’s real median time to one decision. Knolt decides 7× faster than the fastest of the rest, Claude Opus 5.5.

  • Knolt-Talos 1.10.00 sDecided
  • Claude Opus 5.50.00 s
  • Claude Sonnet 5.50.00 s
  • GPT-6.1 Sol0.00 s

Close on paper. Not when it counts.

On incidents caught, the frontier models trail far behind Knolt. Only on correct calls do they come close.

What changed in Security-Bench 1.3 →
Security-Bench 1.3November 2026
69%80% Opus 5.554%74% GPT-6.1 Sol51%72% Sonnet 5.575%83%Knolt-Talos 1.1Incidents caughtCorrect incident calls

Side by side. Even where we lose.

Side by side. Even where we lose.
MeasureKnolt-Talos 1.1KnoltClaude Opus 5.5AnthropicGPT-6.1 SolOpenAIClaude Sonnet 5.5Anthropic
Correct incident calls83%80%74%72%
Incidents caught75%69%54%51%
Alerts that were real88%85%86%86%
Ordinary moments left alone91%89%92%92%
Ranking quality (AUC)91%87%85%84%
Time to one decision0.24 s1.71 s2.81 s1.91 s
Runs on your own GPUYesNoNoNo

Security-Bench 1.3: real security-camera frames and staged security scenes, with incidents and ordinary or lookalike moments. Every model got the same images and the same question.

Measure by measure.

Each scale runs from the pack to Knolt’s score, so the gaps are easy to see.

Correct calls

0%

  • Talos 1.183%
  • Claude Opus 5.580%
  • GPT-6.1 Sol74%
  • Claude Sonnet 5.572%

Scale: 70% to 83%

The most accurate of the four

Incidents caught

0%

  • Talos 1.175%
  • Claude Opus 5.569%
  • GPT-6.1 Sol54%
  • Claude Sonnet 5.551%

Scale: 45% to 75%

The most incidents caught

Ranking quality (AUC)

0%

  • Talos 1.191%
  • Claude Opus 5.587%
  • GPT-6.1 Sol85%
  • Claude Sonnet 5.584%

Scale: 80% to 91%

Closest to a perfect ranking

Alerts that were real

0%

  • Talos 1.188%
  • GPT-6.1 Sol86%
  • Claude Sonnet 5.586%
  • Claude Opus 5.585%

Scale: 80% to 88%

The most alerts that were real

What the test measures.

Security-Bench 1.3 is Knolt’s test of one question: is this a security incident?

It mixes real security-camera frames, taken during robberies and ordinary moments, with short staged scenes filmed the way a security camera sees them: real incidents, and moments that only look like one, such as a hug or a handful of dropped papers.

We count correct calls, incidents caught, alerts that were real, ordinary moments left alone, and AUC: how well each model ranks every incident above every ordinary moment.

  • Robbery
  • Violence
  • Theft and concealment
  • Camera tampering
  • Person down
  • Smoke or fire
  • Ordinary or lookalike