Agent Authorization Trace Audit

• [Promptfoo pricing](https://www.promptfoo.dev/pricing/) makes LLM/agent evaluation, vulnerability scanning, and 10,000 red-team probes per month free. Its enterprise tier adds organization-specific attacks, collaboration, continuous monitoring, compliance dashboards, and services. This validates enterprise budget while pushing the generic red-team price floor to zero.
• [Promptfoo agent red teaming](https://www.promptfoo.dev/docs/red-team/agents/) covers tool discovery, prompt injection, and multiple attack classes. “We also red-team agents” is not differentiation.
• [AgentShield](https://agentshield.ai/), [Torrin](https://torrin.ai/), [Trust3](https://trust3.ai/platform/agent-security/), and [Oktsec](https://www.oktsec.com/) all emphasize real-time tool-call authorization, blocking, identity, audit records, or SIEM integration. Runtime governance is occupied.
• The open [Agent Audit research system](https://arxiv.org/abs/2603.22853) scans agent source code, so repository scanning is also occupied.
• Observability systems and agent SDKs record tool calls, handoffs, and guardrails, but a trace alone does not prove that each side effect remained within the originating user’s authority. The wedge must be **no runtime integration, local processing, reconciliation of existing traces with explicit policy, and an exportable release evidence pack**.
• SERP assessment: candidate terms are likely occupied by OWASP guidance, red-team platforms, security vendors, and editorial content. Do not assume a Top 10 opening. Launch through an open adapter, anonymized cases, and outreach to 30–50 teams shipping tool-using agents.

8h MVP Checklist

1. 1.Define canonical trace, policy, and finding schemas; create 6–8 realistic fixtures with expected findings.
2. 2.Build the generic JSONL parser, schema errors, and local privacy boundary.
3. 3.Build the policy form/YAML import, resource patterns, and approval/reversibility model.
4. 4.Implement deterministic rules, evidence pointers, and the three evidence-status classes.
5. 5.Build the Audit wizard, Results timeline, filters, policy diff, and Markdown/JSON export.
6. 6.Add OpenAI Agents and MCP adapters; reach 20+ regression fixtures.
7. 7.Complete Home, Methodology, Pricing/Pilot, About, FAQ, and Privacy.
8. 8.Add Payment Link/priced lead conversion and content-free analytics events.
9. 9.Run build, unit/E2E, privacy, SEO, mobile, and accessibility checks.
10. 10.After launch, execute the 40-team/150-qualified-visit validation plan before expanding infrastructure.

Don't Build

• ✕Do not expand into active red teaming, runtime firewalls, SIEM, or an enterprise governance platform.
• ✕Do not add a complex backend; the core runs locally in the browser.
• ✕Do not build custom auth, membership, orders, subscriptions, or admin first. Do not omit Payment Link/priced-lead validation.
• ✕Do not send traces, prompts, responses, tokens, credentials, or resource URLs to analytics.
• ✕Do not claim “detects every attack,” “OWASP certified,” “SOC 2 ready,” or any security guarantee.
• ✕Do not support every agent framework for completeness. Guarantee only the three documented formats in v1.
• ✕Do not remove real parsing, policy reconciliation, evidence pointers, redaction, or export to make the MVP lighter.
• ✕Do not let an LLM replace permission rules or label missing evidence as a confirmed attack.