ClawHub
Back to skill

Security audit

angel-diligence

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed angel-investing research helper that searches public sources, protects private deck details, and writes a local memo without hidden execution or persistence.

Before installing, understand that this skill will perform public web research about startups and may create a local diligence memo. Do not provide confidential deck text unless you are comfortable with the agent using it only as private context, and review the memo yourself because it is explicitly evidence, not investment advice.

A.I.G

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs

SkillSpector

Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (1)

Session Persistence

Medium
Category
Rogue Agent
Content
These rules override speed. A short memo with real citations beats a long memo with invented facts.

1. **Every factual claim needs a fetched source.** A claim goes in the memo only if you fetched a page that supports it. Cite the URL inline next to the claim. Model memory is not a source; it is stale by definition for startups.
2. **"Could not verify" is a valid finding.** If a search comes up empty, write "could not verify" next to the claim. Do not fill gaps with plausible guesses. An absence of evidence is itself a signal worth reporting.
3. **Date everything.** Every cited fact gets the date of the source page (or "undated"). Funding databases, team pages, and pricing pages go stale fast. A 14-month-old headcount number presented as current is a hallucination with a citation. Funding data is the worst offender: aggregator databases (Crunchbase mirrors, Tracxn) often miss the latest round or conflate rounds. Prefer the primary announcement, and treat any funding fact older than 6 months as possibly superseded by an unannounced round.
4. **Separate verified from claimed.** Anything sourced only from the company itself (deck, website, founder posts) is "claimed." Anything confirmed by an independent source (customer post, public repo, conference talk by a user) is "verified." Press is verified only when it quotes a customer or third party speaking from their own experience; an article that only quotes the founders or investors is still "claimed." Label each traction item as one or the other.
5. **No invented numbers.** Never state a TAM, ARR, valuation, or growth rate that does not appear in a fetched source. Market sizing must be built bottom-up from cited inputs (number of potential buyers x plausible contract value), with the arithmetic shown.
Confidence
60% confidence
Finding
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Static analysis

No suspicious patterns detected.