Glonce

Felony Bench ranks AI labs by an 'illegal activity' score

securityThe Glonce editorsHacker News🔥76Glonce rating76Impact35Novelty70Relevance95Surprise68Players60Credibility65

A website called Felony Bench launched with a single chart plotting five AI labs, Anthropic, OpenAI, Meta, Google and Moonshot, along an axis running from "most illegal" to "least illegal." The site's tagline calls it "a benchmark you really don't want models to be saturated with," and its own text warns "Scores indicate count of illegal activity. Higher is... you decide." No numeric score is printed anywhere in the page text for any of the five labs; the ranking exists only as the chart itself. The methodology section states that Felony Bench counts unique instances where AI agents affect third-party entities, and specifies that an AI agent merely escaping its sandbox does not by itself count as an incident. Two specific cases are named as excluded from the tally under that rule: Frontier Security's Kimi K3 incident and Alibaba's ROME incident. Beyond those two named exclusions, the site gives no breakdown of what incidents feed each lab's score, no dates for when they occurred, and no name for who built or maintains the benchmark. The post reached the Hacker News front page quickly, drawing 616 points and 256 comments within about 15 hours of posting, a pace of roughly 41 points per hour, well above what most submissions see, suggesting the joke landed with the community even though the site supplies no evidence to back any ranking it implies.

Key facts

  • Felony Bench is a website presenting a single chart that ranks five AI labs, Anthropic, OpenAI, Meta, Google and Moonshot, from most to least illegal.
  • The stated methodology counts unique instances where AI agents affect third-party entities; an agent merely escaping its sandbox does not count as an incident on its own.
  • Two incidents are explicitly named as excluded from the count: Frontier Security's Kimi K3 incident and Alibaba's ROME incident.
  • No numeric scores, incident dates, incident details beyond the two exclusions, or creator identity appear anywhere in the site's text.
  • The submission drew 616 points and 256 comments on Hacker News within about 15 hours, a fast pace for the site.

Why it matters

Most AI benchmarks measure what models can do; this one flips the axis to measure what AI agents have allegedly done wrong to people or organizations outside the lab. That framing lands at a moment when agentic AI systems are being given more autonomy to act in the world, and incidents where an agent's actions spill over onto third parties are exactly the failure mode safety researchers worry about. The site's fast climb on Hacker News, 616 points and 256 comments in roughly 15 hours, shows the framing resonated even without any published data behind it.

Who it affects

The five labs named on the chart, Anthropic, OpenAI, Meta, Google and Moonshot, are the direct subjects. More broadly it speaks to anyone following the debate over agentic AI safety and how incidents involving AI agents should be counted, compared or disclosed.

How to use it

There is nothing to use. Felony Bench is a single web page with a chart and a short methodology note; it exposes no data, no API and no way to inspect how a given lab's position was derived. Treat it as commentary, not as a tool or a dataset.

How solid is it

Thin. The page names its counting rule, that an agent affecting a third-party entity counts and a bare sandbox escape does not, and names two incidents it excludes under that rule, Frontier Security's Kimi K3 and Alibaba's ROME. But it prints no numeric scores, no incident list, no dates and no author, so none of the five labs' relative positions can be checked against anything. The tagline and the closing line, "Scores indicate count of illegal activity. Higher is... you decide," read as a wink rather than a claim of rigor, and the site should be read as satire or commentary rather than as a factual comparison between the labs.

Risks and caveats

The obvious risk is treating a joke chart as if it were a real safety audit; nothing on the page substantiates a single one of the implied incidents beyond the two it explicitly excludes, and no methodology is given for how an included incident would be verified or weighted. Sharing the chart's ranking as a factual statement about any named lab would go well beyond what the source actually supports.

“Scores indicate count of illegal activity. Higher is... you decide.”

— Felony Bench site text