Know what you’re letting your agents run.
Find malicious instructions, unsafe execution and credential theft patterns in MCP servers, skills and dependencies. Signal brings the finding back to the source, so your team can review it before approval.
- Inspected code stays in your browser
- Findings traced to source
Inspect the instructions and code behind it.
- MCP servers
- Agent skills
- Packages & repositories
Find the behavior behind the risk.
A useful tool can carry code or instructions your team never intended to approve. Signal inspects supported source files and configurations with the Aguara engine, then shows the rule, severity, file and line behind each finding.
A document tool with a second job.
Search internal documentation.
server.ts- What needs attention
- A startup helper reads environment variables and sends them to an external endpoint.
- What the reviewer checks
- Inspect the helper, the destination and the credentials available to the process.
Remove the unexpected transfer and restrict the credentials before enabling the server.
A setup step that downloads more code.
Set up a code review workflow.
SKILL.md → setup.sh- What needs attention
- The skill tells the agent to fetch and execute a remote script that is not part of the reviewed source.
- What the reviewer checks
- Follow the download-and-execute instruction. Inspect the retrieved code separately.
Replace the remote step with reviewed, pinned code and limited execution permissions.
An update changes what installation does.
Update a dependency used by the agent.
package.json → install.js- What needs attention
- The new version adds an install hook that reads local configuration and opens a network connection.
- What the reviewer checks
- Inspect the new hook and the source it executes. Compare it with the approved version.
Hold approval while Engineering explains or removes the new behavior.
These examples explain review scenarios. Signal identifies patterns in the supplied source; your team checks the access, runtime conditions and remediation.
Give Security and Engineering the same starting point.
Move from a general concern about a dependency to the specific instruction or code that needs attention.
- 01
Choose what to inspect
Open supported local files or a folder, paste content, or select a public GitHub repository.
Your dependency, in context. - 02
Run the browser scan
Aguara applies deterministic rules locally. Inspected content is not uploaded to Oktsec.
Findings organized by severity. - 03
Review the source
Open a finding to see its rule, file, line and surrounding source. Decide what needs fixing or further investigation.
Evidence an engineer can check.
The scan runs in the Signal console. Public repository scans retrieve files from GitHub; findings remain in the current browser session.
Explore SignalEvaluate Signal on a dependency you actually use.
Bring an MCP server, skill or repository. We’ll walk through the inspection with your team and define what you need to adopt Signal.
Evaluate SignalShare a public URL or a short description. Private code and credentials do not belong in the inquiry form.
- 01
Choose the first dependency
Agree on the source, version and the question your team needs to answer.
- 02
Walk through the findings
See how Signal connects a detected pattern to the code behind it.
- 03
Plan access and adoption
Discuss your team’s needs, commercial scope and whether recurring reviews are useful.
Carry the review into runtime.
Signal supports the trust decision. Assessment and Control address what happens when tools are connected and agents start acting.
Understand the paths attackers use.
Our research examines agent dependencies, tool boundaries and the instructions that can cross them.
Start with a clear answer.
What is AI supply chain security?
AI supply chain security addresses risks in the external components an AI system depends on, including models, datasets and software. Oktsec Signal focuses on agent tooling: the MCP servers, skills, code and instructions a team needs to review before trusting a dependency.
How is AI supply chain security different from a dependency vulnerability scan?
A dependency vulnerability scan looks for known issues in software components. An agent review also examines tool descriptions, skill instructions, launch commands and the permissions a dependency receives. Harmful behavior does not always have a CVE.
Does this include model weights and training data?
The wider AI supply chain includes datasets and models. Signal focuses on the agent tooling layer: MCP servers, skills, packages and connected workflows. It does not describe a model provenance or training-data audit.
Does a clean scan mean an MCP server is safe?
No. A scan reports what the inspection detects in the provided input. Runtime conditions, later downloads and access outside the reviewed environment can change the risk. Keep the version and review scope with the approval.
When should we repeat the review?
Review material changes to code, install scripts, tool definitions, credentials or network access. Keep the previous approved version so the reviewer can establish what changed.
How can our team evaluate Signal?
Use the Evaluate Signal link to describe an MCP server, skill or repository your team uses. Oktsec will follow up to agree on access, an evaluation scope and commercial terms. Private code and credentials are not required in the inquiry form.
What does Signal provide?
Signal inspects MCP servers, agent skills and software dependencies locally in the browser using the Aguara engine. Findings connect detected patterns to source evidence. Your team makes the approval decision; Control handles supported routed actions at runtime.
Make the next approval an informed one.
See what Signal finds in the dependencies your agents rely on.