- Published on
A retrieved document made the model propose a 50,000 euro discount. The prompt injection succeeded, but the action did not. Here is the deterministic policy boundary that stopped it. Part 3 of From Prompt to Proof.
Ai
All Posts
- writings (7)
- reflection (5)
ai (4)
- llms (4)
- engineering (4)
- code (3)
- python (3)
- production (3)
- predictions (2)
- tech-news (2)
- hardware (1)
- paper (1)
- meetup (1)
- feature (1)
- slides (1)
- guide (1)
- audit (1)
- cryptography (1)
- ai-act (1)
- policy-as-code (1)
- opa (1)
- security (1)
- agentic-ai (1)
- pii (1)
- gdpr (1)
- agentic-coding (1)
- software-development (1)
- Published on
Most systems already have logs. Can you prove that the logs you show today are the records the system wrote six months ago? Ed25519 signatures, a SHA 256 chain, canonical JSON, and the one SQL update that broke all of it. Part 2 of From Prompt to Proof.- Published on
Version one of my PII layer masked the words "tax number" and let the Greek AFM through. How I measured Presidio, the checksum trap, and the threshold that silently detects nothing. Part 1 of From Prompt to Proof.- Published on
My real production workflow for shipping features with AI agents: 3 terminals, PID-like feedback loops, independent review, and quality gates. No vibe coding—just controlled engineering that serves real users.