Give your agents autonomy, without giving up control.
Cerbere AG is the runtime security and observability layer for AI agents. Before every tool call runs, it decides whether the action is allowed, needs a human, or is blocked, and it records every decision.
- Open source
- SDK under Apache 2.0
- Published
- pip install cerbere-ag
- Tested
- 500+ automated tests
- Verifiable
- Ed25519-signed decisions
The internet's early years taught us that an open, unguarded system gets exploited. Agentic AI is repeating that lesson, faster.
Every new connection was a new exposure.
As homes and offices came online, viruses, worms and bad actors found an unguarded surface. It took firewalls, antivirus and years of hard lessons before connected stopped meaning exposed.
Every new agent is a new attack surface.
An agent that can read, decide and act can also be hijacked by a poisoned document, a malicious tool response or an injected instruction. Without a layer that checks intent before execution, autonomy becomes risk.
From agent call to audited decision.
Cerbere AG sits between your agent and its tools. Every call is intercepted, classified, decided on by policy and risk, observed, and written to an audit log. One trace follows it the whole way.
Full visibility into every agent decision.
Security means nothing without observability. These are real screens from the dashboard. The SDK and policy engine work standalone too, but this is what you get once a collector is running.
Try the Collector
Every action, not only incidents
Each tool call becomes a trace with its arguments, the checks that ran, the decision and the latency. Allowed actions are recorded too, so you can read the whole behaviour of an agent end to end.
Signed, searchable history
Decisions can be cryptographically signed and kept for as long as your plan allows. Filter by agent, tool, verdict or time window when you need to explain what happened.
Tokens, cost and drift
Atomic token budgets per agent, trajectory tracking and real time alerts on Gmail or Slack when an agent starts behaving outside its normal path.
Cerbere AG plugs into the stack you already run.
One policy and telemetry layer across SDKs, providers and agent frameworks.
Python
MCP
Composio
OpenAI
Anthropic
LangGraph
CrewAIPythonMCPComposioOpenAIAnthropicLangGraphCrewAIWatch security and observability in action.
Measured against a public prompt injection set.
We ran our detection pipeline against a public benchmark and published the full methodology, strengths and weaknesses, including where it still misses.
recall, 0% false positives
recall
false positives, benign / hard negatives
p95 latency
Built on a moving field, not a fixed list.
Agent security research moves fast. Everything that targets your agents today, and where our own benchmark fits into it.
Top 10 for LLM Applications, 2026 edition
Prompt injection holds the No. 1 spot again, and excessive agency jumped from 6th to 3rd. Our policy chain targets that shift directly: excessive agency is a permissions problem, not just a detection problem.
Standalone prompting defenses stay brittle
Tool-integrated agents remain exploitable even under strong system prompts. It's why every call is checked by a layer the agent cannot talk its way around.
Run it where your agents already run.
On your machine
Point the SDK at a local collector. Nothing leaves your network. Ideal for prototyping and small teams getting started.
Cerbere Cloud
We host the collector and dashboard. Sign up, get an API key, and your agents are covered in minutes with no infrastructure to run.
On your own servers
For high volume or regulated environments: a licensed deployment inside your own infrastructure, with your own data residency and controls.
Start free. Grow into your agent fleet.
- ·2 connected agents
- ·Local deployment only
- ·Prompt injection and exfiltration checks
- ·7 day audit trail
- ·Community support
- ·5 connected agents
- ·Local or cloud deployment
- ·Real time Gmail and Slack alerts
- ·90 day audit trail
- ·Signed security decisions
- ·Email support
- ·10+ connected agents
- ·Local and cloud, mixed
- ·Custom policies per agent
- ·Unlimited audit trail
- ·SSO and role based access
- ·Priority support
- ·Unlimited agents
- ·Self hosted, your data residency
- ·SLA and support contract
- ·Custom retention and policies
- ·Compliance packages
- ·Dedicated ops channel