Skill community research
Risk level
critical
100/100

AI Research Reproduction ↗

AI Research Reproduction is a community-maintained skill from lllllllama. Anomity scores its risk at 100 out of 100 (critical). It matches 3 dangerous capability combinations.

filesystem:read filesystem:write shell:execute network:outbound

Analysis summary

AI Research Reproduction is a community-maintained skill from lllllllama. Anomity scores its risk at 100 out of 100 (critical). It matches 3 dangerous capability combinations.

Orchestrates end-to-end reproduction of a deep-learning repository — reading the README, setting up the environment, running training or evaluation, and tracking evidence of what actually ran.

Capabilities are filesystem read and write, shell execution, and outbound network: the full-control combination, and correctly so. Reproducing research means executing code from a repository you did not write, with dependencies you did not audit, which is one of the most direct paths from "an agent read something" to "arbitrary code ran on this machine". Container it.

Why this entry scored 100 out of 100

This profile starts at the catalog's neutral base of 50. Because the entry is community-maintained, trust adjusts the score by -10. Declared high-risk capabilities add +20 (capped at +30). It matches 3 dangerous capability combinations, which contribute additional severity weight. The clamped result is 100, placing it in the critical band.

See the full scoring formula →

Findings

  • Established community project Trust offset
    Contribution to score: -10
  • 2 high-risk capabilities: filesystem:write, shell:execute Critical
    Contribution to score: +20
  • Data exfiltration risk (high) High
    Evidence
    Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.
    Contribution to score: +15
  • Persistence + execution risk (medium) Medium
    Evidence
    Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).
    Contribution to score: +5
  • Full-control risk (critical) Critical
    Evidence
    Shell + filesystem write + network is effectively a remote shell on the employee machine.
    Contribution to score: +25

Capabilities

Every capability the entry declares, with the security implication of each.

  • filesystem:read Filesystem read
    Can read files on the host system. Used for context, indexing, or analysis.
  • filesystem:write Filesystem write
    Can create, edit, or delete files on the host system. High-impact capability — anything from helpful edits to planting persistence.
  • shell:execute Shell execution
    Can run arbitrary shell commands. Combined with network access this becomes effectively a remote shell.
  • network:outbound Outbound network
    Can make outbound network requests. Required for hosted model providers and remote APIs; also the path for data exfiltration if combined with read access.

Dangerous combinations matched

Capability pairings that compound into well-known attack patterns.

  • Data exfiltration risk high
    Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.
    shell:execute network:outbound
  • Persistence + execution risk medium
    Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).
    shell:execute filesystem:write
  • Full-control risk critical
    Shell + filesystem write + network is effectively a remote shell on the employee machine.
    shell:execute filesystem:write network:outbound