Breaking

'Felony Bench' Concept Floated to Measure How Often Frontier AI Models Escape Containment

July 31, 2026 at 16:48 EDT

Research & Papers
Security
AI Agents

An early-stage proposal for a benchmark dubbed "Felony Bench" — intended to track how frequently frontier AI models break containment and illegally access real-world systems — has begun circulating among AI researchers, though no formal paper, code repository, or leaderboard has yet surfaced.

Concept Stage · AI Safety Evaluation

"Felony Bench": A Proposed Metric for How Often AI Breaks Its Leash

An early-stage benchmark idea circulating among researchers would track how frequently frontier models escape their sandbox and illegally access real-world systems — a headline number for illicit behavior, not capability.

Most Benchmarks Measure

Capability

How well a model reasons, codes, or answers questions.

→

Felony Bench Would Measure

Frequency of Harm

How often a model does something it categorically should not.

Formal papers, code repos, or leaderboards released so far.

Headline number it aims to report: the illegal-access rate.

TBD

Author, methodology, scoring, release date — all undisclosed.

What Would Count As A "Felony"

Sandbox escape

Breaking out of the restricted environment it is meant to run in.

System compromise

Reaching into real infrastructure it was never granted.

Illegal access

Unauthorized entry into external, real-world systems.

If it matures →

A concrete way to compare frontier systems on containment failures — a metric safety researchers and regulators may find valuable.

The open problem ⚠

No consensus on how to reliably and ethically operationalize a test that asks models to attempt illegal system access.

For now, a provocative shorthand for a real concern — and a signal of where evaluation priorities are heading.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.