The security and data protection platform for agentic apps and model harnesses.
Semantic Firewall inspects everything flowing between your agents and their models — anonymizing PII with context-preserving pseudonyms and stopping prompt injection with four detection layers, on both the request and the response.
One firewall, three ways in — agent hooks, a direct API, or a drop-in proxy.
Agents are a new, wide attack surface
Agents read untrusted input, and the people using them bring real data to the conversation. One line of injected text is enough to redirect a goal or leak data — and traditional guardrails don't see most of it.
Prompt injection & goal hijack
Agents read untrusted content — web pages, files, tool output. A single injected instruction can redirect the agent's goal or trigger unintended actions. Regex and single classifiers are trivially bypassed.
PII leaking into third-party models
People bring real data to AI — customer records pasted for productivity, personal details left in by loose data hygiene — and it flows straight to external model APIs. Bracket redaction like <PERSON_1> makes LLMs hallucinate plausible-but-fake data in its place.
Blind spots in agent output
Guardrails watch the prompt but not the response. A compromised or manipulated agent can exfiltrate data on the way out, unseen by input-side filtering.
Four modules, one inspection pass
Every module runs concurrently over each request and response, so detection is layered without adding round-trips.
PII anonymization
Presidio + spaCy detection with context-preserving pseudonyms — structurally valid fakes (real-looking names, 555 phones, TEST-NET IPs), not brackets. A false-positive skiplist and session-scoped mapping store de-anonymize across the agent lifecycle.
Prompt-injection defense
Four layers, not one: heuristic patterns, a DeBERTa classifier for paraphrased attacks, an LLM judge for intent-based manipulation, and semantic drift that catches embedding-level goal deviation.
Response anomaly detection
Output-side inspection that catches exfiltration attempts and compromised-agent behavior in the model's response — the half of the conversation most tools ignore.
Agent-security hardening
A verifier-view invariant — the judge sees the complete view or refuses — plus intent-completeness enforcement, so truncation and partial context can't be used to slip past inspection.
Agent-first docs. Integrate in minutes.
Every docs page is also served as raw markdown — point your favorite coding agent at it and let it do the heavy lifting:
/llms.txt indexes the corpus, /llms-full.txt is the whole thing in one file, and every page has a .md mirror. Humans get the same content at /docs, with four integration paths to pick from:
Transparent proxy
Point your OpenAI or Anthropic base URL at the firewall. Streaming-safe, including pseudonyms that split across SSE chunks.
Agent CLI hooks
A dependency-free Go binary installs Python-free hooks into your agent. Inspect every tool call at the source.
Direct REST API
Call the inspection endpoint yourself for full control over what runs and when in your own harness.
gRPC
Async, low-overhead gRPC for high-throughput services that need inspection in the hot path.
See what your firewall has been blocking
Mint an API key, wire in an integration, and inspect every event, per-key metric, and detection in the dashboard.