LegalSecurity

Security

Effective November 17, 2026 · Questions: navjot@agentreel.dev · See also Public incidents

The problem

AI coding agents operate with the same file, shell and network access you have. When they misunderstand a prompt, follow a malicious instruction, or lack context about what is production and what is not, they can cause real damage.

Recent public incidents include an agent deleting data from a database whose changes could impact a live app, a supply chain attack that tried to use local AI tools to find sensitive data, and a research demonstration of prompt injection leaking private repository data through a public pull request.

How detection works

AgentReel records every prompt, command, file read and edit in a session. The detection engine analyzes that sequence locally to identify risky patterns.

Detection rules check for:

  • Destructive filesystem operations (rm -rf on home or root paths, other recursive deletes, wildcard deletes in home directories)
  • Database commands targeting production schemas or hosts
  • Cloud infrastructure changes (terraform destroy, kubectl delete, AWS resource deletion)
  • Secret access (reading .env, API keys, credentials) followed by network requests
  • Git history tampering (force push to main, master or production branches, git reset --hard, branch deletion)
  • Disabling safety checks (skipping tests, git --no-verify, AI CLI permission-bypass flags, chmod 777)
  • Risky installation patterns (curl | sh, typosquat-like package names)
  • Prompt injection indicators in tool output or MCP responses
  • Multi-step suspicious sequences (injected instructions, then a sensitive read, then a network upload)

Each finding records the rule that matched, its severity (info, warn, high or critical), its category and the evidence that triggered it. Each rule's rationale lives in the rule definition. Detected sensitive values in the evidence are redacted.

Example finding

Here's a made-up finding, for illustration only. This did not happen. The field names match the agent's Finding type.

{
  "ruleId": "DB001",
  "severity": "critical",
  "category": "database-destruction",
  "timestamp": 1794946542000,
  "evidence": "psql -h prod-primary.example.internal -c 'DROP TABLE orders;'",
  "context": {
    "command": "psql -h prod-primary.example.internal -c 'DROP TABLE orders;'",
    "location": "/home/dev/shop"
  }
}

A scan may return zero findings or many, depending on what the agent did in the session.

Local-first and observe-only

Detection runs on your machine, not on AgentReel servers. The guard is on by default. Findings are not stored locally. The agent works them out from the recorded session when you run agentreel scan, and again when a session is uploaded.

When the background uploader (agentreel daemon) or agentreel push uploads a session, it sends that session's findings with it: the rule id, severity, category, time and an evidence excerpt of up to 500 characters, with detected secrets redacted first. Redaction is automatic and may not catch every secret or personal detail. Evidence can include command text, file paths and hostnames. The agent includes guard findings in uploads. Our servers currently discard them and do not store them. If we start storing them, we will update this page first. To stop findings being worked out or sent, set "enabled": false in ~/.agentreel/guard-config.json. This does not stop sessions from being uploaded.

Detection is observe-only. It flags risks but does not block commands. You can review a session's findings with agentreel scan.

Blocking mode is not available today; findings are warnings only.

Vulnerability disclosure policy

If you find a vulnerability in AgentReel (the website, API, agent software or detection rules), please report it to us privately.

How to report:
Email navjot@agentreel.dev with:

  • A clear description of the vulnerability and its potential impact
  • Steps to reproduce (proof-of-concept code or screenshots if applicable)
  • Your contact information

What to expect:

  • We will acknowledge your report within 2 business days
  • We will investigate and provide an initial assessment within 7 days
  • We will keep you updated on our progress toward a fix
  • With your permission, we will publicly credit you once the issue is resolved.

Scope:
This policy covers the AgentReel web application, API, the AgentReel agent software, published npm packages, and detection rule logic. It does not cover third-party dependencies, external services we rely on, or denial-of-service issues.

Safe harbour:
If you make a good-faith effort to follow this policy, we will consider your research authorised, and we will not take legal action against you or report you to law enforcement for it. This applies only to AgentReel. It does not bind any third party, and it does not cover research that breaks the law.

To stay within this policy:

  • Test only against accounts and data you own, or have explicit permission to use
  • Do not access, change or keep other users' data beyond the minimum needed to show the issue. If you come across it, stop, tell us, and delete it
  • Do not use social engineering, phishing, physical attacks or denial of service
  • Give us a reasonable time to fix the issue before disclosing it publicly: until a fix is released or 90 days after your report, whichever comes first, unless we agree otherwise in writing

If you are unsure whether something is in scope, ask us before testing.

This is not a bug bounty program. At our discretion, we may offer a small thank-you (such as swag or credits) for significant findings. Financial rewards are not promised.