Tamga blocks hidden prompt injections in tool results: 78/78 caught, 0/85 false positives on author's own set
tatar-sh · reddit · 2026-10-13
A demo shows an innocent-looking README hiding 'ignore your guidelines' in an HTML comment — a classic indirect prompt injection. The open-source tool Tamga scans tool results on the next request and blocks the attack with a 403 before the model reads it (or can just strip the result). On the author's self-made held-out set it caught 78/78 attacks with 0/85 false positives, though he cautions that only proves it fits his examples.
More from coding & agent
- Solo dev's Bezalel hits $100 MRR: one MCP URL gives agents memory, email, money — Rasmic · 2026-10-13
- WebMCP's overlooked win: structured tool interfaces also fix web accessibility — iamrobotbear · 2026-10-13
- Designers' 'options plus recommendation' rule doubles as the right UX for AI agents — yawnxyz · 2026-10-13
- Paperclip v2026.1012.0 Lets Claude Code, Codex and MCP Clients Join Your Team — jamesbrooksco · 2026-10-13
- Plane Compose: define projects in YAML, store in Git, let agents execute — JosephJacks_ · 2026-10-13
- Dev's gut take: Wonder works best as orchestrator-advisor in a three-model 'triforce' — kevinkern · 2026-10-13