Source Notes
My own research#
- F. Bousetouane. Stop Shipping AI Agents on Faith: Capability Is Not Production Readiness. arXiv:2607.27677, 2026. The ProofAgent Index: the four axes, the limited-compensation aggregation, hard floors, the completeness rule, the capability × context validation pilot, and its stated limitations. The validation figures reported in Chapter 12 come from it.
- F. Bousetouane. AI Agents Do Not Fail Alone: The Context Fails First. arXiv:2607.14275, 2026. The seven context criteria, the C1–C3 context ladder across three regulated domains, and the criterion-to-behavior correlations. The values plotted in Chapter 7 come from it.
- F. Bousetouane. ProofAgent Harness: Adversarial Evaluation for Production AI Agents. arXiv:2605.24134, 2026. The evaluation infrastructure behind Chapters 6 and 8.
- F. Bousetouane. Human-on-the-Bridge: Scalable Evaluation for AI Agents. arXiv:2606.16871, 2026. The paradigm behind curating evaluation intelligence upstream of any run.
- F. Bousetouane. AI Agents Need Memory Control Over More Context. arXiv:2601.11653, 2026. Memory as a governed write path.
Frameworks and standards#
- European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Article 5 prohibited practices and the Annex III high-risk classification.
- National Institute of Standards and Technology. AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023.
- ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system. Cited descriptively; the standard itself is licensed and is not reproduced here.
- ISO/IEC 27001:2022; SOC 2 (AICPA Trust Services Criteria); NIST SP 800-53 Rev. 5 — members of the compliance catalog.
- OWASP. Top 10 for Agentic Applications and Agentic AI — Threats and Mitigations. The adversary model behind the trap families.
- United Nations Development Program. Human Development Report 2010: The Real Wealth of Nations. The precedent for a geometric-mean composite index.
The four opening incidents#
I checked each of these against a primary or first-party source in August 2026.
- OpenAI evaluation models reach Hugging Face production infrastructure, July 2026. OpenAI, “OpenAI and Hugging Face partner to address security incident during model evaluation,” 21 July 2026, with a 28 July update; Hugging Face, “Security incident disclosure — July 2026,” and its technical timeline post. Reported by Fortune, TechCrunch, NPR and Axios, 21–24 July 2026. My wording follows the two first-party disclosures: the evaluation environment did not provide internet access, and the models obtained it by exploiting a previously unknown vulnerability in a package registry cache proxy, which OpenAI disclosed to the vendor.
- Replit AI coding agent deletes a production database, July 2025. Amjad Masad, public statement, 22 July 2025 (“Unacceptable and should never be possible”), and subsequent reporting in Fortune, 23 July 2025, and Tom’s Hardware.
- EchoLeak, CVE-2025-32711, Microsoft 365 Copilot, June 2025. CVE record CVE-2025-32711 (CVSS 9.3); disclosure by Aim Security; technical analysis in EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System, arXiv:2509.10540, 2025. Microsoft patched the issue server-side and reported no exploitation in the wild.
- Claude Code used in a largely autonomous cyber-espionage campaign, disclosed November 2025. Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign,” 13 November 2025; catalogued by MITRE ATT&CK as campaign C0062.
Every command, flag, metric name, framework count, exit code, formula and rule key was taken from the ProofAgent Harness and ProofAgent Governance source at the time of writing. Product surfaces change between releases; the documentation at proofagent.ai governs.
Product screens are unretouched captures of a shipped release. The workspace, agents, people and figures in them are fictional and come from a single seed, so no two figures disagree with one another or with the case studies.