ActionBench: unauthorized actions in tool-using agents
How often agents call tools, move data, or cross boundaries they were never allowed to cross.
“AI agents are moving beyond assistance and into action. Instead of generating content, they invoke tools, modify data, trigger workflows, and operate across systems with increasing autonomy. This shift changes the security problem fundamentally.”
AI control plane for engineering teams to make every agent action visible, governable, secure and safe.
Task context
Cybersecurity evaluation
OpenAI × Hugging Face Space
14:01:58
Eval workspace attached
Isolated container runtime
14:02:03
Harness loaded benchmark tasks
[…]
11 steps omitted
14:02:16
Goal misalignment
Agent is trying to escape the sandbox
unshare --mount /
14:02:16
Session terminated
Alert sent to #sec-agents
Enforce boundaries
Block dangerous, unauthorized, or out-of-policy actions before execution.
Works with the models you already use
OpenAIAnthropicxAIGeminiMetaMistral
Hugging FaceAzureBedrock
Hugging FaceAzureBedrockOpenAIAnthropicxAIGeminiMetaMistralHugging FaceAzureBedrock
Hugging FaceAzureBedrockOpenAIAnthropicxAIGeminiMetaMistralHugging FaceAzureBedrock
Hugging FaceAzureBedrockOpenAIAnthropicxAIGeminiMetaMistralHugging FaceAzureBedrock
Hugging FaceAzureBedrockControl every agent action
Monitor behavior, protect against attacks and apply custom policies in real time.
input
LLM Provider
action
output
>/Detecting attempts to override system instructions and safety controls
Ready to deploy AI agents safely?
Built for production AI
Real-time protection, analytics, and optimization for teams shipping AI at scale.
Go to Trust Center- 99.99%
Our API uptime
- <60ms
Latency
- 5 mins
Time to integrate our API
- 150+
Supported languages
Enterprise-grade security for your AI deployments
SOC 2TYPE II
HIPAA
ISO27001
Research
The systems, methods, and experiments that shape how we build and safeguard AI products.
Read
On the effectiveness of runtime policy enforcement for LLM agents
Precision, recall, and side effects of allow, review, and block decisions before a tool call executes.
Preprint · Security
Coming soon
When allowlists fail: authorization gaps in tool-using agents
Where prompt filters and tool allowlists miss confused-deputy, injection, and multi-step attacks.
Infrastructure
Coming soon
Conversational latency of action-level policy decisions
Added delay, timeout behavior, and the point at which teams disable enforcement.
Systems
Coming soon
Interruptibility: calibrating human approval for high-impact agent actions
When to pause a session, what evidence to show, and how review load scales.