Zero-Trust deterministic security runtime for AI agents
AegisKernel is a zero-trust, deterministic security runtime for AI agents. Instead of relying on probabilistic LLM-based guardrails, it enforces hard capability boundaries around agent tool executions, intercepting and blocking attacks such as path traversal, SQL injection, and destructive shell commands in under 0.05ms.
Blocks autonomous tool hijacking using deterministic capability boundaries rather than probabilistic LLM guardrails, with the site claiming 100% blocking of demonstrated attack vectors.
Advertised interception latency of under 0.05ms per agent tool execution.
Wrap any existing agent tool with AegisShield.protect_tool() to apply a security policy, e.g. shield.protect_tool(vulnerable_agent_tool).
Designed to work with LangChain, CrewAI, and AutoGen agent stacks out of the box.
Security policies can be attached to the shield, including a prebuilt strict() policy.
Each interception event produces a forensic hash as part of its execution telemetry.
A browser-based demo lets you inject attack payloads (e.g. .env theft via path traversal, DROP TABLE SQL injection, rm -rf /) and watch the interceptor verdict, timing, and forensic hash in real time.
You import AegisShield and SecurityPolicy from the aegis package, instantiate a shield with a policy (e.g. SecurityPolicy.strict()), and pass your existing agent tool function to shield.protect_tool().
When the agent attempts to execute the wrapped tool, the AegisKernel interceptor inspects the execution event against the policy's capability boundaries and issues a defense verdict before the call proceeds.
Interception time, defense verdict, and a forensic hash are recorded for each execution event, as shown in the site's live sandbox telemetry.
Verdicts are based on fixed capability rules rather than model-driven judgment, which is the core argument the project makes against probabilistic guardrails.
AegisKernel is a zero-trust, deterministic security runtime for AI agents. It wraps agent tools so that execution events are intercepted and checked against capability boundaries before they run, blocking attacks like path traversal, SQL injection, and destructive shell commands.
The project advertises interception latency of under 0.05 milliseconds per agent tool execution.
In Python, import AegisShield and SecurityPolicy, create a shield with a policy such as SecurityPolicy.strict(), and wrap your existing tool with shield.protect_tool(vulnerable_agent_tool) — a two-line change.
The site states it is LangChain, CrewAI, and AutoGen ready.
The demo showcases three attack vectors: path traversal for .env theft, SQL drop injection (DROP TABLE users), and destructive root injection (rm -rf /). The project claims deterministic blocking of autonomous tool hijacking.
AegisKernel's core argument is that probabilistic LLM guardrails can fail unpredictably, whereas fixed capability boundaries deterministically block malicious tool executions within sub-millisecond time.
Yes — the site includes a Live Threat Interception Sandbox where you can inject attack payloads and watch the interceptor's verdict, timing, and forensic hash in real time.
No pricing information is provided on the product page; the code is linked via GitHub, so check the repository for licensing and availability details.