What is Aethra AI?
Aethra AI is a runtime defence and governance platform for AI agents, built by a pre-seed startup. At its core is an enforcement layer that tracks where every piece of data in an agent's world came from — trusted or untrusted — and stops untrusted data from driving actions that could leak, change, or destroy sensitive systems. The architecture follows a Planner / Quarantined Extractor / Executor model aligned with dual-LLM security research (CaMeL). The open-source SDK (Apache 2.0, Python and Node/TS) is available on GitHub, with an optional hosted control plane for approvals, audit evidence, and kill-switch functionality.
Who it's for
- Security engineers — need a deployable boundary that blocks prompt-injection attacks without relying on pattern matching or cloud round-trips.
- Platform teams running agentic systems — want provenance tracking and tool-level policy enforcement that works offline inside their own stack.
- Compliance and risk officers — require auditable evidence of what agents actually did, including approvals, blocked actions, and outbound destinations.
Key features
Provenance-boundary enforcement
Every value an agent handles is labelled trusted or untrusted based on its origin. That label travels with the data through the workflow. When an action would let untrusted data reach a dangerous capability — send, write, call out, execute — the runtime blocks it or holds it for a human, and records the decision in terms of sources, not attack phrasing.
Open-source runtime with offline operation
The core protection runs on the machine executing the agent. Blocking decisions do not depend on Aethra's servers being up. The SDK is available for Python (aethra-ai) and Node.js / TypeScript (@aethra-ai/sdk), both implementing the same enforcement model.
Optional operator control plane
When connected, the control plane gives your team visibility into enforcement events, approval workflows, agent registration and suspension, and compliance evidence views — all built from what the runtime actually recorded. The block decision stays local; the cloud never becomes a black box.
Tool-level policy actions
Per tool policy, the runtime can allow a call, block it, or require human approval. Blocks and holds are logged with the tool name and tainted source identifiers — never raw prompt content or secrets.
What stands out
Aethra AI is not a filter — it's a provenance boundary that decides from data origin, not from matching attack phrases.
That distinction matters. Most defences try to recognise malicious instructions by wording, which fails against novel or obfuscated attacks. Aethra instead labels data by where it came from and enforces at the tool boundary, so untrusted content simply cannot drive dangerous actions regardless of how it's phrased. In benchmark testing on AgentDojo, Aethra blocked 35/35 injection goals with 0 false positives. The company is currently raising a €1.0–1.5M pre-seed round.
Worth checking out if…
You run AI agents that touch sensitive systems and want a defence that works offline, ships as open source, and gives your security team evidence — not guesswork — when something goes wrong. If you need org-wide visibility, approvals, and kill-switch control, the optional control plane is the natural next step once the open-source core is in place.
