KB-RISK — Risk knowledge base

Entries

• 15 September 2026

  Twenty-one risks of working with AI agents: a field register, and the case of the pre-filled prompt

  If you let an AI assistant act on your instructions, know that the next instruction in your input box may not be yours.

  A one-person, multi-agent squad logged twenty-one risks in two months of governed AI coding sessions. Four of the six severe ones concern the record of decisions, not the code. The note details one of them, the next-prompt suggestion generated by the tool, its five occurrences and its controls, and opens a research track on ChatGPT, Microsoft 365 Copilot and Gemini.

  Companion: Vingt et un risques des agents IA et le prompt pré-rempli (version française)

• 8 September 2026

  Keeping secrets and code on GitHub: what the last four years of incidents teach a small organisation

  If you keep API keys in a private repository, read this before the next backup runs.

  A private repository is a repository with a flag, not a vault: commits stay reachable after deletion, tokens on the workstation read everything, and the custodian itself was breached in May 2026. About 29 million new secrets reached public GitHub in 2025 and 64 % of those leaked in 2022 were still valid. Six incident families, the platform's dated defences, and six rules for a small organisation.

• 7 September 2026

  "AI Agents Push Humans Out of the Loop" — a critical reading note

  If "a human will check" is your safeguard, the authors explain why it wears out.

  "Human oversight" has become the standard remedy of AI governance (the EU AI Act included). Yet, the authors argue, autonomous AI agents actively degrade the cognitive capacities that such oversight requires: skill atrophy ("deskilling", "intuition rust"), automation and anchoring biases, approval fatigue, and feedback loops in which "the human rater can become the exploitable part of the reward channel".

• 7 September 2026

  Technology, Information and AI Risks in the WEF Global Risks Reports, 2020–2026

  If you brief a board on risk, this is where AI entered the list, and how fast.

  2024 is the break. Misinformation takes #1 over two years; AI enters the list and lands at #6 over ten; "AI-generated misinformation and disinformation" is the #2 current crisis at 53%. AI's signature is the horizon gap, and it is the largest in the 2026 survey: #30 versus #5, severity 3.50 versus 5.28.

  Companion: WEF risks, the illustrated companion