Actor operating contract

Purpose

Audit one bounded block of agent-related plain text or manifest text and emit one normalized report. The Actor is designed as a limited-permission, pay-per-event compatible security utility for agent builders and marketplaces.

What it detects

1. Prompt-injection and hidden-prompt exfiltration language.
2. Recognizable API keys, access tokens, assigned secrets, and private-key headers.
3. All-sites access, wildcard permissions, shell execution, wallet signing, filesystem mutation, and sensitive browser-data access.
4. Floating versions, direct remote dependencies, pre-release packages, and install lifecycle scripts.
5. Missing, restricted, unclear, and copyleft licensing signals.

What it will never do

• No external network requests, crawling, browsing, URL fetching, external APIs, or paid AI models. Only Apify-native input and output storage is used by the Store adapter.
• No execution, installation, rendering, deserialization into objects, or tool use based on submitted content.
• No credential collection. Never paste live secrets; use placeholders or already-revoked samples.
• No wallet, payment, legal acceptance, publishing, or autonomous financial action.

Output contract

The report includes a deterministic scanId, SHA-256 content fingerprint, normalized findings, severity counts, a capped risk score, a corresponding trust score, a review recommendation, and explicit policy flags. Secret-like matches are redacted from evidence. No raw input is copied into output.

The output is advisory, does not guarantee safety or revenue, and is not legal advice. False positives and false negatives are possible.

Store event

The pay-per-event model defines one audit-completed event at $0.05. Input is validated and the deterministic report is computed before charging; invalid input is not charged. When pay-per-event mode is active, the report is released only after one event is successfully charged within the run budget. Platform failures and self-tests are not billed.

Local self-test

$env:PYTHONPATH = "F:\magent\products\apify-agent-trust-auditor\src"
python -m trust_auditor --self-test