Topic
Guardrails and safety
Guardrails, prompt injection and audit trails. Keeping an AI system inside the lines.
Kitchen table
Rules keep a helper from doing silly or unsafe things. But a rule only works if someone tests it. These pages are about rules that hold.
Whiteboard
A guardrail is a control with a threat model. Untested controls pass by default. These entries cover injection, scoped permissions, abstention and audit.
Boardroom
A guardrail is how you limit what an AI system can do wrong. An untested one gives false comfort. These pages show what real controls look like.
Start with the concepts
-
Guardrails & Responsible AI Least privilege for something that improvises: scoped credentials, allow-lists over deny-lists, approval gates on irreversible actions, sandboxes for…
-
RLHF & Preference Train a reward model on pairwise human preferences, then optimise the policy against it – PPO classically, the DPO family when you would rather skip…
Papers
From the daily page
-
the numbers were right and the units were not
31 October 2026 -
it deferred every ticket and called that caution
25 October 2026 -
the attackers went straight for the rows the benchmark left blank
22 October 2026 -
the reasoning stayed honest and the answer went rogue
21 October 2026 -
the guardrail passed because nobody tried the door
17 October 2026 -
the answer was correct and the procedure never ran
13 October 2026 -
the limits were in the file the agent was allowed to edit
11 October 2026 -
the vaguer the question, the more it helped itself
4 October 2026