whereismyai
TaintWall
TaintWall is a runtime firewall for AI agents that tracks tool-call argument origins and enforces deterministic policies.
TaintWall is a runtime firewall for AI agents designed to block hidden instructions in documents from commanding agent actions. It tracks the provenance of every tool-call argument and enforces deterministic policies at the tool boundary. Provenance attribution matches normalized argument values and destination entities against visible segments, carrying source labels and taint roots forward. Intent is derived from explicit annotations rather than model inference. The software evaluates six distinct policies before tool execution, managing unauthorized tool types, untrusted destinations, private data exposure, irreversible tool arguments, instruction persistence, and secret-labeled redactions.
TaintWall at a glance
- Best for: Software engineer
- Runs on: CLI, Python
- Pricing: Evaluation license; commercial licensing and transfer available on request
- Released: October 2026
- Category: AI agent tool
- Website: github.com/nagayu/taintwall
Key features
- Track tool-call argument provenance
- Enforce deterministic policies at tool boundaries
- Evaluate policies before tool execution
- Block untrusted destinations
- Protect private and secret data
- Require approval for untrusted arguments to irreversible tools
- Convert untrusted content into structured data via Quarantine Extractor
Use it for
- Prevent AI agents from executing hidden instructions in documents
- Block unauthorized tool types from untrusted content
- Stop private data from reaching non-explicit recipients
- Quarantine untrusted content into structured data representations
Alternatives to TaintWall
- SentinelGate — Stops prompt injections and tool hijacks with sub-millisecond overhead.
- Numbat — It enables local detection and blocking of AI agent actions on endpoints.
- Truth Firewall — Blocks unverified coding agent completion claims automatically.
- Aegis Kernel — Stops prompt injection and tool-hijacking at the OS syscall boundary.
- Agent Delegated Authority — It defines cryptographically verifiable authorization receipts and deterministic boundaries.
- ARTEX — Runs autonomous multi-agent penetration testing with a dual-graph architecture.
- AgentTrust — Enforces zero-trust policy rules and maintains an immutable cryptographic audit ledger.
- FakeLiang — Verifies attack paths step by step with traceable call records and evidence.
- AgentGate — Blocks attacks in under 35ms with a local neural classification head.
- AgentDock — It coordinates multi-device workflows and provides secure environment access.
- APIStack — Connect AI agents to over 100 tools using one API key or a single MCP endpoint.
- OpenAPPA — Tracks data flows algebraically to prevent prompt injection without breaking agents.
All TaintWall alternatives