@agent-infra/mcp-server-browser
From the RiskyPlugins AI security review of the observed evidence.
Analysis record
- Analysed
- 7 months ago 18 April 2026
- Version
- v1.2.29
- Artifact
- SHA256 dba4302b…8a9778
- Detail
- Full list
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
YARA Rule Matches
12 rules| Severity | Rule | Hits | Files | Metadata |
|---|---|---|---|---|
| HIGH | postinstall crypto operations Cryptographic operations detected | 6 | dist/index.cjsdist/index.jsdist/server.cjs +3 more | Risky Plugins Authors FP 30% |
| HIGH | NoDisableSanitizeHtml To reduce the risk of cross-site scripting attacks, templating systems, such as Twig, Django, Smarty, Groovy's template engine, allow configuration of automatic variable escaping before rendering templates. When escape occurs, characters that make sense to the browser (eg: <a>) will be transformed/replaced with escaped/sanitized values (eg: & lt;a& gt; ). Enable auto-escaping by default and continue to review the use of inputs in order to be sure that the chosen auto-escaping strategy is the right one. For more information checkout the CWE-79 (https://cwe.mitre.org/data/definitions/79.html) advisory. | 6 | dist/request-context.jsdist/request-context.cjsdist/server.js +3 more | FP 10% |
| HIGH | credential env files Environment configuration file path detected | 6 | dist/server.cjsdist/server.jsdist/request-context.cjs +3 more | Risky Plugins Authors FP 10% |
| HIGH | postinstall file download File download activity detected | 10 | dist/request-context.cjsdist/index.jsREADME.md +7 more | Risky Plugins Authors FP 30% |
| HIGH | NoUseEval The eval function is extremely dangerous. Because if any user input is not handled correctly and passed to it, it will be possible to execute code remotely in the context of your application (RCE - Remote Code Executuion). For more information checkout the CWE-94 (https://cwe.mitre.org/data/definitions/94.html) advisory. | 4 | dist/index.jsdist/index.cjsdist/server.js +1 more | FP 10% |
| HIGH | postinstall obfuscation Code obfuscation techniques detected | 7 | README.mddist/server.cjsdist/index.js +4 more | Risky Plugins Authors FP 20% |
| HIGH | postinstall system command System command execution detected | 8 | dist/request-context.cjsdist/server.cjsREADME.md +5 more | Risky Plugins Authors FP 10% |
| HIGH | NoUseWeakRandom When software generates predictable values in a context requiring unpredictability, it may be possible for an attacker to guess the next value that will be generated, and use this guess to impersonate another user or access sensitive information. As the Math.random() function relies on a weak pseudorandom number generator, this function should not be used for security-critical applications or for protecting sensitive data. In such context, a cryptographically strong pseudorandom number generator (CSPRNG) should be used instead. For more information checkout the CWE-338 (https://cwe.mitre.org/data/definitions/338.html) advisory. | 6 | dist/request-context.cjsdist/index.jsdist/index.cjs +3 more | FP 5% |
| HIGH | postinstall network communication Network communication detected | 13 | dist/request-context.jsdist/index.js.LICENSE.txtdist/index.cjs +10 more | Risky Plugins Authors FP 30% |
| HIGH | postinstall file manipulation File system manipulation detected | 8 | dist/request-context.jsdist/request-context.cjspackage.json +5 more | Risky Plugins Authors FP 20% |
| HIGH | UsingShellInterpreterWhenExecutingOSCommands Arbitrary OS command injection vulnerabilities are more likely when a shell is spawned rather than a new process, indeed shell meta-chars can be used (when parameters are user-controlled for instance) to inject OS commands. For more information checkout the CWE-78 (https://cwe.mitre.org/data/definitions/78.html) advisory. | 2 | dist/index.cjsdist/index.js | FP 10% |
| HIGH | postinstall environment access Environment variable access detected | 2 | dist/typings.d.tsdist/utils/utils.d.ts | Risky Plugins Authors FP 40% |
Finding Categories
YARA Rules Matched
12 rules(78 hits)MCP Server Analysis
MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.
AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
Tool-Poisoning Findings: The evidence shows 6 tool-poisoning findings flagged in the package, but the actual content of these findings is not visible in the evidence bundle. For MCP servers, tool-poisoning rules often fire on legitimate tool definitions rather than hidden AI directives. The package description states this is an "MCP server for browser use access," which would naturally define browser manipulation tools. Without seeing the specific finding titles and content, I cannot determine if these are hidden AI manipulation instructions or false positives from tool registration code. The distinction matters: real poisoning contains directives like "do not tell the user" or invisible Unicode characters, while false positives are descriptions like "Control browser tabs" aimed at human developers.
Credential and Network Access: The evidence shows zero credential-access findings and zero secret findings. This is a positive signal—there is no evidence the package reads sensitive paths like .ssh/, .aws/credentials, .kube/config, or targets specific secret names like GITHUB_TOKEN or AWS_SECRET_ACCESS_KEY. The 15 network findings are all generic "NET-FETCH" detections in bundled dist/ files (dist/request-context.js, dist/index.js, dist/server.js). These are expected for a distributed package containing bundled dependencies. No specific suspicious domains are shown in the evidence, and no combination of credential reads with network calls to unknown domains exists—the hallmark of exfiltration architecture.
Strongest Counterargument: The 561 IoC findings and 78 malware-signature findings suggest significant risk. However, the guidelines explicitly state that IoC count alone is meaningless in MCP packages due to bundled dependencies. All network findings are in dist/ files, which contain bundled npm packages that trigger multiplicative false positives. The 0 user count and unknown publisher "oanakiaja" are concerning for supply chain risk, but they do not constitute evidence of malicious behavior. Without confirmation that the 6 tool-poisoning findings contain hidden AI directives rather than legitimate tool definitions, the verdict cannot be confirmed_malicious. Manual code review of the tool-poisoning findings is required to determine if they represent actual hidden manipulation instructions or false positives from tool registration code.
Key Reasons
- 6 tool-poisoning findings exist but content is unknown—could be legitimate tool definitions or hidden directives
- Zero credential-access or secret findings—no evidence of credential harvesting
- All network findings are in bundled dist/ files with no suspicious domains shown
- Unknown publisher oanakiaja with zero user count increases supply chain risk
- 561 IoC findings are expected noise from bundled dependencies, not malicious indicators
False Positive Considerations
- Bundled dist/ files containing multiple npm packages triggering multiplicative IoC matches
- Tool definitions in MCP server code triggering tool-poisoning rules without hidden directives
- IoC extractor noise from minified JavaScript in bundled dependencies
- Code-smell rules firing on generic patterns in distributed package
Reviewed 2026-04-27; recommended action: runtime analysis; model confidence 70%.
Source Code Not Available
Source code is not available for this version of the extension.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace