Where Are All the Prompt Injection Damages?
joshua_saxe · x · 2026-08-23
After evangelizing about prompt injection risks, the author observed minimal real-world damages compared to overall cyber losses. Analysis suggests universal jailbreaks may have become harder, especially for Anthropic and GPT-5.6 Sol models. However, goal hijacking remains practical. The article warns that while fixing global security tech debt is expensive, agents could make exploiting these vulnerabilities significantly cheaper for attackers.
More from Safety
- Tension between data center opposition and AI industry expansion — NathanpmYoung · 2026-08-23
- GovAI Hiring Entrepreneurs-in-Residence with $150k Funding for AI Governance Projects — Manderljung · 2026-08-23
- AI must identify itself even if it passes the Turing test — arieljalali · 2026-08-23
- AI-generated reviews create risks for tech decision-making — DavidLinthicum · 2026-08-23
- Research: Misconfigured Admin Prompts Can Invert LLM Safety Layers — Simple_Passion_7741 · 2026-08-23
- Qwen Model Generates 60 Prompt Injections, Security Tools Fail to Block — JLeonsarmiento · 2026-08-23