Skip to content

getting-started-security-reviewer.md

Latest commit

 

History

History
53 lines (38 loc) · 2.04 KB

getting-started-security-reviewer.md

File metadata and controls

53 lines (38 loc) · 2.04 KB
Raw

Getting Started — Security Reviewer

Last reviewed: 2026-06-06 Persona: Security, compliance, or audit reviewer evaluating TeaAgent

Start with the trust model, not the feature list. TeaAgent's value is making agent actions provable — not claiming zero risk.

Primary artifact

Trust and Audit Whitepaper — guarantees, non-goals, failure modes, verification commands.

Enterprise NIST mapping detail: Security Whitepaper.

Evidence to request from operators

  1. Sample run audit log: .teaagent/runs/<run_id>.jsonl
  2. Chain verification output: teaagent audit verify --root .
  3. Compliance export: teaagent audit export --audit-log <run_id> --output bundle.json
  4. Config lint: teaagent doctor config-lint --root .
  5. Permission policy: teaagent approval list --root .

Key controls to test

Control Test
Read-only blocks writes Run with --permission-mode read-only; confirm ToolPermissionError
Approval exactness Scoped grant must match tool + arguments (see tests/test_approval_token_exactness.py)
Path containment Symlink / traversal blocked (tests/test_ws3_schema_path_containment.py)
Compliance mode TEAAGENT_COMPLIANCE_MODE=1 aborts on audit disk failure
Prompt injection boundaries prompt-injection-trust-boundaries.md

Known gaps (honest)

  • No SOC 2 certification bundled with the OSS harness
  • Provider-side retention is out of scope
  • Remote MCP servers are high-trust-boundary — require explicit review
  • Plugin code is not cryptographically signed by default

Automated regression suite

python3 -m pytest tests/test_ws3_compliance_audit.py \
  tests/test_ws3_strict_audit_chain.py \
  tests/test_prompt_injection_boundaries.py \
  tests/test_approval_token_exactness.py -q

When to reject TeaAgent

See When Not to Use TeaAgent for deployment shapes where another product is a better fit.