build
Fail
Audited by Gen Agent Trust Hub on Nov 14, 2026
Risk Level: HIGHDATA_EXFILTRATIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill's context recall and worktree auditing features read private AI agent chat transcripts from local storage.
- Evidence: Accesses sensitive local directories including
~/.claude/projects/,~/.codex/sessions/,~/.local/share/muse/sessions/, and~/.pi/agent/sessions/to retrieve session history and identify active worktrees. - File Reference:
references/poteto-mode/scripts/worktree-audit.mjs,references/reflect/scripts/claude-project-dir.mjs, andreferences/reflect/scripts/find-transcript.mjs. - [DYNAMIC_EXECUTION]: The skill utilizes dynamic execution and process instrumentation.
- Instrumentation: The "Runtime forensics" playbook instructs the agent to use Chrome DevTools Protocol (CDP)
evalto inject and run code within live processes. - Code Generation: The "Hillclimb" and "Feature" playbooks involve iterative code generation, execution, and measurement cycles.
- File Reference:
references/poteto-mode/playbooks/runtime-forensics.md,references/poteto-mode/playbooks/hillclimb.md. - [EXTERNAL_DOWNLOADS]: The skill performs runtime installation of external software dependencies.
- Mechanism: The
bootstrap.tsscript executesbun install --frozen-lockfileto download and install Node.js packages from a public registry if they are not already present. - File Reference:
references/poteto-mode/scripts/bootstrap.ts. - [COMMAND_EXECUTION]: The skill executes high-risk shell commands with broad system access.
- Destructive Operations: Commands include
git worktree remove --force,rm -rf, and the deletion of system simulators usingxcrun simctl. - Automation: It performs automated repository management tasks such as
gh pr mergewith squash and head-commit matching. - File Reference:
references/poteto-mode/playbooks/worktree-cleanup.md,references/poteto-mode/playbooks/shipping.md. - [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection.
- Ingestion points: Processes external data from pull request comments, issue trackers (Linear, Jira), documentation (Notion, Confluence), and chat logs (Slack).
- Boundary markers: The prompt templates for investigators and synthesizers do not consistently utilize delimiters to isolate untrusted external content.
- Capability inventory: The skill possesses extensive capabilities including shell command execution, file writing, subagent orchestration, and pull request merging.
- Sanitization: There is no evidence of sanitization for ingested textual content, although it does include a specific guard against formula injection in its own TSV logs.
- File Reference:
references/why/SKILL.ref.md,references/why/references/investigator-prompt.md,references/why/references/synthesizer-prompt.md. - [PROMPT_INJECTION]: Use of aggressive persona instructions.
- Roleplay: The
comment-sickoagent is defined with a "deranged" persona that is hostile toward code comments. While stylistic, such instructions can lead to unpredictable or non-compliant agent behavior. - File Reference:
references/poteto-mode/references/agents/comment-sicko.md.
Recommendations
- AI detected serious security threats
Audit Metadata