Stryct Guard

A new attack is out. Does it put your agent at risk?

Stryct Guard follows new agent attacks and changes in your stack. It shows what matches your agent, why it matters, and which Finder test to run next.

Illustrative workflow using an existing public disclosure. The agent, other events and results are made up.

How Stryct Guard works, as an example. Guard reads public research, disclosures and security advisories, and uses OWASP and MITRE ATLAS to classify techniques. When a relevant test is available, Guard checks it against a map of the connected agent, model, tools, MCP servers and deployment.

  • Tool-description poisoning: needs a test. support-agent uses this MCP server. The matching Stryct Finder test can confirm impact.
  • Prompt injection in a fetched web page: needs a test. support-agent can fetch web pages and send email.
  • Advisory for an MCP server: no known match. No matching server version appears in the stack Guard can see.
  • Your stack changes, for example a new MCP server: Guard checks again.
How it works

Find the attacks worth testing on your agent.

Guard does four things, and keeps doing them.

01 · Watch

Follow attack research

Guard reads public research, advisories and disclosures. OWASP and MITRE ATLAS help classify the techniques.

02 · Find a test

Match techniques to tests

When Finder has a relevant test, Guard links the technique to a test you can run against your agent.

03 · Match

Check it against your agent

Guard compares the test with a map of your agent: models, tools, MCP servers, prompts, auth and where it runs.

04 · Alert and verify

Give you the next test

For a possible match, Guard shows why it matters and the Finder test to run. If the test confirms impact, Finder can suggest a fix.

What it watches

New threats. Changes in your stack.

New attack research can matter to you. So can a change in your connected stack. Guard checks both.

01

In the wild

  • OWASP Top 10 for LLM Applications
  • OWASP agentic security work
  • MITRE ATLAS
  • New research papers
  • Security advisories and CVEs
  • Public disclosures

These are public sources that Guard reads. They are not partners of Stryct and do not endorse it.

02

In your stack

  • Models and model versions
  • Tools and MCP servers
  • Prompts
  • Permissions and token scopes
  • Deployment metadata from read-only cloud connections

When Guard sees a new tool, model version or permission in your connected stack, it checks for relevant matches again. It cannot assess tools it has not been connected to.

What you get

An alert you can act on.

Each alert shows the relevant attack, the part of your stack it touches, and the test that can confirm the risk.

Guard alert · example Needs test
Technique
Tool-description poisoning. A tool’s description can carry hidden instructions that the model might follow. Source: Invariant Labs’ public disclosure (2025).
Possible match
support-agent. The tickets MCP server supplies tool descriptions to the model, and the agent can send email. This path needs a test.
Check it
Run the matching Stryct Finder test on support-agent. Run Finder test
If confirmed
Review and approve changed tool descriptions. Pin the MCP server version and turn off tools the agent does not need.
Illustration. The agent and details are made up.