Open Source Security for AI Agents
AgentGate intercepts every agent action before execution — verifying identity, validating delegation chains, and detecting behavioral drift in real time.
Request Early AccessView on GitHub
or
$npm install agentgate-pdp
Open SourcePython SDKTypeScript SDKLangGraph ReadyKill Chain DetectionMITRE ATLAS Mapped
Enterprises are deploying autonomous AI agents at scale — but the security infrastructure hasn't kept up. Every agent is a potential attack surface.
Traditional identity systems grant access once and assume good behavior. They cannot detect when an agent exceeds its delegated scope mid-task.
When Agent A delegates to Agent B delegates to Agent C — who authorized the final action? No existing tool answers this.
An agent's behavior shifts silently over time. By the time you notice, the damage is done.
Every agent action is scored across four dimensions before it's allowed to run. No agent bypasses the gate.
25%
Ed25519 JWT tokens with scope embedded in the signed credential — immutable after issuance, offline-verifiable with the public key. No database lookup required.
25%
Full chain traversal at every authorization call: every ancestor's scope is verified. Atomic revoke_chain neutralizes an agent and all descendants in one call.
30%
Embedding-based semantic scoring: action + resource (85% weight) vs. declared purpose. Justification is capped at 15% — cannot be used to bypass a misaligned action.
20%
Per-agent velocity baselines with trust decay over time. Dormancy followed by sudden high-volume activity is itself a risk signal — no static thresholds.
Beyond single-request
Each individual request may look clean. AgentGate examines the full 5-minute sequence. Bulk reads followed by an export. A read followed by a delete on the same resource. Progressive sensitivity escalation. Directory sweeps across 6+ prefixes. Patterns that only become visible across multiple calls — and that no rule-based system can catch.
DENY
BULK_READ_THEN_EXFIL
10 reads → export
DENY
READ_THEN_DELETE
Read → delete same file
ESCALATE
SENSITIVITY_RAMP
Low → CRITICAL escalation
ESCALATE
DIRECTORY_SWEEP
6+ prefix recon
1
Drop in your API key — one line of code
2
See every agent action in real time — attacks blocked live
Demo scenario — AgentGate intercepting a simulated multi-agent attack sequence in real time
Full Demo — 5 min
Watch a live run — real agents, real attacks, real-time blocking.
The regulatory and threat landscape is converging. Enterprises need answers now.
OWASP LLM06
Excessive Agency — agents granted permissions beyond their declared scope, acting outside their intended purpose. Listed as a critical risk in OWASP Top 10 for LLM Applications.
OWASP Top 10 for LLM Applications, 2025
MITRE ATLAS
Adversarial ML tactics against AI systems now formally catalogued — reconnaissance, privilege escalation, and data exfiltration all apply to autonomous agents.
MITRE ATLAS, 2024
August 2026
EU AI Act high-risk obligations take effect — enterprises have months, not years, to implement governance controls for high-risk AI systems.
EU AI Act (Regulation 2024/1689)
Regulatory pressure and adversarial sophistication are converging. Teams without agent governance controls today face compliance exposure by Q4 2026.
Drop-in integration. No framework changes. No rewrites.
Python 3.10+TypeScript / Node.jsLangChainLangGraphAutoGenMCP
We're onboarding select enterprise pilot teams with limited availability.
Priority given to teams running LangGraph, LangChain, or custom agent frameworks in production with real compliance requirements.
Dedicated onboarding
1:1 setup with the founding team
Pilot pricing
Flexible pricing for early adopters
Direct influence
Shape the roadmap with your use case
We'll review your request and get back to you within 48 hours.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。