Status register · evidence
Technical Evidence
What is implemented, in development, and planned — kept in one register and rendered on every page. Statuses are conservative: implemented means built and tested; limited means implemented with small-scale tests or known gaps. Only we promote a status, and each promotion needs dated evidence.
Register last verified: 2026-11-09. This page is the single home of published limits and of the one number below.
Runtime Security
Screening of untrusted text by channel
User input, retrieved documents, tool results, email — verdict with reasons.
Prompt-injection and malicious-context detection
Direct and indirect injection in inspected text.
Action gate
Human confirmation when an outbound action's destination appears only in untrusted content.
Tool-call monitoring and policy enforcement
Beyond the action gate.
MCP and tool security
Model Context Protocol sessions and tool integrations.
Agent memory integrity and memory-write protection
Long-term memory against poisoning and tampering.
Passive pilot (offline replay + written summary)
Recorded traffic replayed offline with our engineering team.
Live passive mode in the customer's system
Passive monitoring inside the customer's live environment.
Signed, verifiable action receipts
Cryptographically signed receipts for allowed and blocked actions.
AI Security Audit
Training Data Security
Ranked review queue of suspicious samples.
Backdoor-trigger and label-manipulation detection
Scope confirmed against the evidence on record.
Template Security — static analysis
Chat templates in GGUF and safetensors packages.
Template Security — behavioral mode
Behavioral testing beyond static analysis.
Security Intelligence Layer
Event ingestion, normalization, correlation, provenance, cross-product insights
One view of AI security events across the lifecycle.
Trust
Privacy-conscious logging
Fingerprints where possible, published retention schedule.
Configurable failure policy for external model errors
Fail-closed available as an option; not the default.
Non-root service execution
All services as unprivileged users.
Access control and data minimization
Least-privilege access, minimal stored data.
Want these statuses tested against your own agent? A limited, non-invasive evaluation with our engineering team.
Request a Technical Evaluation