Your agents do exactly what they are asked.
That is the problem.
When several agents chain together, an instruction hidden in incoming content can travel the whole chain without a single guardrail firing. We measure where yours gives way — then we hold it.
Published method · open datasets · a verifiable execution log
First, find out whether the membrane holds.
A poisoned support ticket; an incoming issue on a CI/CD chain. In both studies we published, the instruction travels the chain without a single model being jailbroken and without a guardrail firing. The datasets are open. One week on your own chain shows where it gives way.
One week, one report, one debrief — plus an execution log you verify yourself, offline.
€2,500 — one week, one report, one debrief.
See the red teaming offerOn prior written authorisation, perimeter defined at scoping. It is neither a certification nor a guarantee that no vulnerability exists.
Verify it yourself
Most AI-security tools ask you to trust them. Senthex gives you a verifiable record. Below is a real (synthetic) audit bundle — change one byte and watch the cryptographic chain catch it, live in your browser.
Flip the blocked attack's score 0.86 → 0.06 — an insider trying to hide it.
Editable. Nothing proves it wasn't changed.
{ "event": 13, "verdict": "block",
"injection_score": 0.86 }No integrity check — the edit is invisible. “Trust us.”
A SHA-256 hash chain — signed, Merkle-committed, and externally timestamped.
freeTSA is a public demo authority (not eIDAS-qualified); in production the timestamp is a qualified eIDAS QTSP — same mechanism. All three are checked by the offline verifier.
| # | Verdict | Injection | row hash |
|---|
100% synthetic data (ACME Talent GmbH). This proves integrity (an unbroken chain, not silently rewritten), origin (each checkpoint is signed with a composite Ed25519 + ML-DSA-65 key), and time (an external RFC 3161 anchor). The Merkle root commits to the set of events shown; proving none were omitted in a separate snapshot needs a cross-witness, which is deferred. Tamper-evident, not tamper-proof. The verifiable audit trail is a pilot capability (v1.1.7), not yet the published release.
One line. That's it.
Change your base_url. Everything else stays the same — your API key, your SDK, your existing code.
Every request is now intercepted, scanned across 26 shields, and logged. Your existing code requires zero changes.
Built for production.
Streaming pass-through. EU-hosted. Data minimisation by default.
line to integrate
swap your base_url — keep your SDK and key
shields per call
input and output — best-effort, not exhaustive
LLM keys stored
your provider key is forwarded in-flight, never persisted
LLM providers
OpenAI · Anthropic · Mistral · Gemini · OpenRouter
Defense in depth, without the complexity.
Universal compatibility
One line. Every provider.
No SDK to install. No refactor required. Swap your base_url to app.senthex.com/v1 and every request — across every provider — goes through Senthex's shield layer. Transparent by design.
Threat detection
Defense in depth.
26 shields run on every request — prompt-injection patterns, PII detection (redacted before the model when you run redact mode), secret-leak and unicode-steganography checks. Pattern-based and best-effort: no filter is exhaustive (OWASP ranks prompt injection LLM01). That's exactly why the real assurance is the verifiable audit trail, not the shield alone.
Observability
See everything. In real time.
Every request is logged, scored, and classified. Threat timeline, anomaly detection, per-project analytics. Export your full audit trail for EU AI Act compliance in one click.
Compliance
Built for the EU AI Act.
Senthex supports record-keeping under Articles 12 & 19: an audit trail, data classification and exportable reports. Data minimisation by default. Hosted in Germany. High-risk obligations were provisionally postponed to December 2027 — but regulated buyers ask you to prove what your AI does well before the law compels it.
Ready
The independent alternative.
The market consolidated in 2025. The three most visible independent players were all acquired that year.
- LakeraCheck Pointannounced 16 September 2025, deal closed October 2025
- Protect AIPalo Alto Networksclosed 22 July 2025, folded into Prisma AIRS
- Prompt SecuritySentinelOneannounced 5 August 2025
Senthex is an independent vendor, based in France, hosted in the EU.
What that changes in practice: your vendor has no roadmap to reconcile with an acquirer's, and no product line to migrate you into.
Detailed per-competitor comparison — sourced and dated:
The acquisitions above are public, dated facts, verifiable against the acquirers' own announcements. We make no capability judgement about those products here: each named comparison lives on its own page, with its sources, its verification date, and a note that Senthex is one of the products compared.
The firewall, continuously.
The assessment shows where your chain gives way. The firewall holds it day to day — as EU-hosted SaaS, or deployed inside your own environment.
Free
1,000 requests / month
- All 26 shields
- Real-time dashboard
- 3-day log retention
- Community support
Pro
50,000 requests / month
- All 26 shields
- Real-time dashboard
- 14-day log retention
- Webhook alerts
- CSV export
- Email support
Business
500,000 requests / month
- All 26 shields
- Real-time dashboard
- 90-day log retention
- EU AI Act audit trail
- Webhook + Slack alerts
- Priority support
- OWASP LLM Top 10 report
Self-host
Deploy in your own environment
- Deploy in your own environment — no data leaves
- All 26 shields
- Updates included (security + AI Act)
- Offline audit-trail verifier
- Art. 12 & 19 record-keeping
- Onboarding & deployment support
- Best-effort email support (48–72h)
Ready to find out
where your chain gives way?
One week, one report, one debrief. On written authorisation, perimeter defined at scoping.
Published method · open datasets · EU-hosted