Tamga proxy blocks hidden prompt injections in files read by AI agents

A developer released Tamga, a self-hosted proxy that sits between apps or agents and the LLM provider to catch indirect prompt injection in tool results. It detects hidden HTML comments, invisible Unicode characters and encodings, then either blocks the request or strips the poisoned fragment. The code is open under AGPL-3.0.

• Tamga parses each part of a request separately and reports which role a finding came from
• STRIP mode removes the poisoned tool result without ending the agent session
• Median scan time is 0.23 ms per chat-sized prompt (p95 0.84 ms)
• On the held-out set it caught 78 of 78 attacks with 0 false positives on 85 benign texts