AI agents install packages without human review — so we check every package and its source (registry, URL, publisher, name originality) and hand the agent a deterministic verdict: SAFE / UPGRADE / AVOID — exactly at install time.
Traditional scanners (Snyk, Aikido, Trivy) run in CI — after the code is committed. But the agent's decision happens earlier, on the developer machine, at npm install time. That moment is the attack surface.
North Korean campaign specifically designed to bypass AI coding agents — malicious npm/PyPI packages that read like legitimate docs and pass agent "reasoning". Humans would spot it; the agent installs it in seconds.
attack pattern: agent-assisted supply chain compromiseMalware distributed through package READMEs and metadata containing prompt-injection text aimed at the agent's context window — the agent itself becomes the delivery vector.
prompt injection → agent executes attacker instructionsSigil's 2026 scan of 39,000+ packages built for AI agents found malicious patterns in 4.2%. Typosquats, credential stealers, and install-time payloads — invisible to post-commit scanners.
sigilsec.ai — State of AI Agent Supply Chain Security 2026Reference MCP servers shipped by major AI labs carried critical vulnerabilities. If the tool layer itself can't be trusted blindly, the verdict layer must be deterministic and signed — not another AI guess.
Aikido AutoTriage = AI heuristics → non-deterministicVerdicts right at the agent's decision point — at install time and in every PR. Plus a provenance check of every source: registry, repository URL, publisher, name originality.
The agent checks whether the vulnerable code is actually reachable in your codebase — per framework, not just per package.
Malicious packages often look legitimate — the attack is in the name, the URL or the registry, not the code. We verify provenance at install time:
A package can be clean code and still arrive through a malicious door — we flag that before the install completes, and the whole lockfile is validated in one call (check_install_plan).
No known vulnerabilitiesTested against continuously updated vulnerability and malware intelligence. The agent confirms relevance in your code.
Vulnerability, but a fix existsWe tell you exactly which version to upgrade to — and the agent checks what it breaks.
Malware, typosquat, active exploitKnown malicious packages, lookalike names, actively exploited vulnerabilities. Do not use.
A deterministic pipeline that neither a language model nor a malicious package README can influence.
Vulnerability and malware databases, merged into one consistent view with full provenance — refreshed automatically, always current.
Versioned rules compute the verdict. Blue = authoritative engine output. Reproducible and auditable — same input, same verdict, no exceptions.
The agent queries the verdict via MCP and adds relevance: reachability, upgrade impact. Purple = said by the agent. It never rewrites the verdict to "safe".
A language model never touches your security decisions. The verdict is computed by deterministic, versioned rules, and every decision has a full audit trail — reproducible end to end, certifiable for audits.
The AI agent can only lower the relevance of a verdict — "not reachable in our code" — it can never rewrite it to "safe". Prompt injection from a package README never reaches the verdict, because the engine does not read it.
For the public sector, critical infrastructure and regulated industries — no US CLOUD Act exposure, GDPR-compliant from day one. All hosting and data processing is EU-only.
Vibe coding means your agent writes code and installs dependencies at machine speed — without a human in the loop. These are the risks that come with it, and what we do about each:
An agent installs what it is told to — a typosquat or lookalike name (axiοs vs axios) passes "looks right" reasoning in seconds. Soverenia checks every package name for originality before install.
Malicious instructions hidden in a package README or description target the agent's context window — the agent itself becomes the delivery vector. Our engine never reads package content, so injected text can't reach the verdict.
→ verdict computed from metadata only, never from package textAn unofficial registry, a dead or redirected repository, a sudden maintainer change or a license switch between versions — clean code can arrive through a malicious door. We verify provenance: registry, URL, publisher history, lockfile integrity.
→ source & URL checks on every install (see above)A dependency deep in the tree carries a known CVE — you didn't pick it, your agent did. Soverenia flags it with a verdict and tells you exactly which version fixes it, checked against continuously updated threat intelligence.
→ SAFE / UPGRADE / AVOID per package, with upgrade targetFree distribution via the open-source Agent Guard, paid platform for teams and sovereign enterprise deployments.
Short answers to the questions we hear most.
Those tools scan your code or images in CI — after the commit. Soverenia sits at the agent's decision point, at npm install time on the developer machine, and hands the agent a deterministic verdict via MCP. It's not a broader scanner; it's the missing layer exactly where agents act.
No. The verdict is computed by deterministic, versioned rules — reproducible and auditable. The AI agent calling us can only add local context (e.g. "not reachable in our codebase") — it can never rewrite a verdict to SAFE.
npm, PyPI, Maven, Go, crates.io, NuGet and RubyGems — full scope today, with continuously updated vulnerability and malware intelligence behind every verdict.
EU-only hosting, no US CLOUD Act exposure — a natural fit for NIS2 / CRA / EU AI Act compliance. We don't store your queries: this website sets no cookies and runs no analytics or third-party trackers (fonts are self-hosted; the only localStorage entry is your theme preference, which never leaves your browser). The only data kept is basic server access logs (IP, timestamp, requested path) for security and rate limiting, deleted after 30 days.
Yes — the Enterprise plan includes a local-hosted, sovereign EU-only deployment with audit export and SLA. The CLI (Agent Guard) is open source and free.
Yes — the Agent Guard CLI and MCP server are free and open source, with verdicts for public repositories included. Paid plans (see Pricing above) add private repos, CI gates, policy profiles and sovereign enterprise deployments.
Local-hosted sovereign deployment (EU-only), audit export, SLA, NIS2/CRA reporting — tell us about your setup and we'll get back within one business day.