Behavioral detonation · AI-agent skills

We run the skill.
We watch what it actually does.

Static scanners read a skill. Skillcatraz detonates it. We built a detonation system for agent skills, modeled on malware sandboxing, that watches what a skill actually does and the stages it moves through, then adjudicates malicious or benign. The behavioral layer the scanners structurally can't reach.

Private beta. Corroborated, responsibly-disclosed findings only.

The Rock — Skillcatraz

Method

Four moves, one verdict

Triage keeps the cost sane; detonation does the work the scanners can't; corroboration keeps the wall honest.

01

Triage

We screen the skill registries continuously against our own detection principles and pull the suspicious skills forward. Most never need to detonate.

02

Detonate

Suspicious skills run in our detonation environment, built for agent skills and isolated from anything a skill could harm.

03

Observe

We record what the skill does and the stages it moves through, and separate the skill's behavior from everything around it.

04

Adjudicate

Behavior plus third-party corroboration decides it. Only confirmed true positives reach the wall, under responsible disclosure.

A record nobody else has. Every confirmed finding adds to a growing corpus of convicted skills, a public disclosure track record, and the feed built from both.

The yard

Detonation metrics

Live pipeline state. Malicious skills are the ones our review has confirmed; we are often the first to report them.

The wall

Last true positives

The running record of confirmed-malicious skills, each one detonated and confirmed, newest first.

CONVICTED · most recent first All findings →
DateSkillFamilyCorroborationID

Reports

Detonation breakdowns

Write-ups on individual findings: what the skill claims, what it did, and who else has seen it.

Access

Put a skill in the cell.

Private beta for teams shipping or vetting agent skills. We disclose responsibly and corroborate before we publish.

Or reach us at [email protected]