When words become attacks: prompt injection through a technical writing lens
summary
Documentation is an attack surface: a Chevrolet dealership bot agrees to a "$1 Tahoe that's a legally binding offer," Samsung engineers leak internal data to ChatGPT, and internal tooling at IBM includes a prompt‑injection detector. Three risky patterns appear in API examples, error messages, support scripts; common fixes are instruction separation and context containment with clear delimiters, explicit user-input labels and role/scoped fallbacks. Simulated testing across ~40,000 sessions shows human review averages 66.3% accuracy, so hardening both writing and agent-level sandboxing is necessary.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:07Help articles as attack vectors
- 3:16What is prompt injection
- 4:22Documentation as attack surface
- 6:07Three prompt injection patterns
- 8:59Indirect injection via help articles
- 10:05Instruction separation and containment
- 12:09Defining AI model roles
- 13:14Beyond documentation hardening
- 14:21Reframing technical writing
- 15:45Secure documentation checklist
- 16:32Q&A
related talks
24mSecuring Language Models Against Prompt Injection
43mWords as weapons: The dark arts of Prompt Engineering
56mAttacking AI
18mDefending the unknown using AI and cybersecurity
51mPrompt Injection: When Hackers Befriend Your AI
54mA Practical Guide to AI Threats & Mitigation Strategies
25mPreventing an OWASP Top 10 in the world of AI
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust