VirusTotal opened an official OAuth MCP path for Claude Code at https://ai.virustotal.com/mcp, letting analysts ask coding agents to look up file hashes, URLs, domains, and IP reports without pasting a personal VirusTotal API key into mcp.json. If you run Claude Code, Codex, or similar MCP clients on investigation laptops, that changes what your agent can reach over the network and what indicators leave the machine when someone says “check this hash.”
The integration is not a silent antivirus layer. VirusTotal’s VTAI overview states plainly that a missing report, pending analysis, or zero detections is not a safety guarantee. The current Claude Code plugin guide (release 0.1.6 in the raw doc) further narrows risk: no upload hook ships in the plugin anymore—agents are steered toward get_file_report by SHA-256 inside host file permissions, while explicit submissions need deliberate user intent and can publish samples to VirusTotal’s community.
| Question | Short answer |
|---|---|
| Am I affected? | Yes, if anyone on your team installs the virustotal@virustotal plugin or adds the HTTP MCP server on a machine that holds downloads, incident logs, or malware samples |
| Is there a fix? | No CVE—risk is process and policy. Approve the connector, train analysts on hash-only workflows, and block surprise URL submissions |
| Does data leave my machine? | Hash lookups send the hash; URL/domain/IP queries send the full indicator; submissions can upload bytes or public download links |
| What changed recently? | Plugin 0.1.6 removed the Node upload hook documented for 0.1.5 and earlier; updates require claude plugin marketplace update virustotal |
| What should I do today? | Inventory MCP servers, pin the plugin version, and read access.md quotas before agents automate lookups |
What VTAI adds to an agent harness
VTAI (VirusTotal AI) is VirusTotal’s agent-first API surface: MCP tools plus a Bearer-token REST API at https://ai.virustotal.com, with free tiers inside quotas and no classic VT API key required for OAuth onboarding.
The documented MCP tool set is small and investigation-shaped:
get_file_report— lookup by hash; does not upload the fileget_url_report,get_domain_report,get_ip_report— network indicatorssubmit_file,get_submission,get_analysis— contribute or resume analysis when explicitly requested
Supported clients listed in November 2026 docs include Claude Code, Codex (HTTP or stdio), and Agy (Antigravity CLI), plus plugins such as VT Sentinel, OpenClaw, and Hermes for hash and reputation tooling.
For Claude Code specifically, VirusTotal publishes a first-party install path on Connect Claude Code to VirusTotal MCP:
claude plugin marketplace add VirusTotal/virustotal-mcp
claude plugin install virustotal@virustotal
After /reload-plugins, analysts authenticate from /mcp with browser OAuth (Google sign-in flow in VirusTotal’s connection guide). A manual HTTP alternative exists:
claude mcp add --transport http virustotal https://ai.virustotal.com/mcp
…but VirusTotal warns against duplicate server entries if the plugin is already installed. The plugin also exposes a skill entry point documented as /virustotal:threat-intelligence.
Attack surface: what the agent can leak
Treat VTAI like any high-privilege MCP server—it is legitimate threat intel, not a toy feed.
Hash lookups vs uploads
The 0.1.6 plugin documentation is explicit that the remote connector cannot read local paths. The expected workflow:
- User authorizes checking a download.
- Claude computes SHA-256 within host-permitted file scope (Claude Code’s own permissions—not VTAI’s).
- Claude calls
get_file_reportwith that hash.
A missing report does not authorize executing the file or uploading it. That discipline matters because coding agents otherwise “helpfully” run unknown binaries after a conversational nudge.
Explicit uploads follow a different, higher-risk path: submit_attachment with a public HTTPS download link (up to 24,000,000 decoded bytes per VTAI overview) or directing the user to the VirusTotal web uploader when no safe link exists. Standard submissions are shared with VirusTotal’s community and security partners per the Claude Code guide—so private documents, internal source trees, credentials, and personal data require human permission, not model enthusiasm.
URL queries are full-fidelity exfiltration
The Claude Code doc’s URL guidance is easy to miss in a long session: URL lookups disclose the complete URL, including query parameters and fragments. If an analyst pastes a signed S3 URL, password-reset link, or internal ticket URL into chat, the agent can ship that string to VTAI when it “looks up the report.” Prefer get_domain_report when a domain reputation check is enough.
This is adjacent to the MCP tool-poisoning and rug-pull patterns we document for less reputable servers—except here the capabilities are official and documented. The defense is the same class: know what each tool sends, and deny writes you did not request.
Quotas, receipts, and “try again” loops
VirusTotal’s access.md quota model is unforgiving for autonomous agents:
| Behavior | Quota impact (per VTAI docs) |
|---|---|
| Any explicit report or analysis read | Consumes query quota, including cache hits, repeats, and not_found |
| New file contribution confirmed unknown | Uses contribution limits, not report-query quota |
| Returning an existing file after upload | Still costs one report query |
Contribution limits are 20 files per fixed minute and 500 per UTC day per Agent Token identity or OAuth account, shared across upload paths. VirusTotal explicitly warns not to register duplicate identities to evade limits—exactly what an over-eager agent might attempt if prompts say “keep trying until clean.”
Analysts should also preserve get_submission receipts on the same MCP connection that performed a submit; migrating from a manual claude mcp add entry to the plugin requires reading VirusTotal’s migration section before deleting the old server.
Codex, hosted Claude, and other clients
VTAI is not Claude-only. The VTAI overview lists Codex over HTTP or stdio and documents Bearer-token REST at https://ai.virustotal.com/api/v3/… for headless workflows. VirusTotal’s access.md also records ChatGPT Work OAuth validation (September 2026) for hosted connectors—separate from Claude Code’s plugin path.
If your team runs multiple harnesses, treat VTAI agent tokens and OAuth connections as one identity per person: registration creates one credential per agent in VTAI’s model, and revoking it disables MCP, REST, and stdio together. Inventory which laptops hold ~/.config/vt-mcp/token files before you rotate incident responders between machines.
Cyber Security News on November 16, 2026 highlighted the analyst workflow angle—bringing detection counts, analysis dates, and source links into conversation—but trade summaries may lag plugin 0.1.6 (no local upload hook). Defer to VirusTotal’s own caution that AI narration is not a verdict. Your runbook should require click-through to the VT report or --json excerpts archived in the ticket, not paraphrase alone.
Plugin 0.1.5 → 0.1.6: upload hook removal
If your team installed VTAI earlier in 2026, the breaking narrative is the retired local upload hook. VirusTotal’s doc tells you to:
claude plugin marketplace update virustotal
claude plugin update virustotal@virustotal
…then /reload-plugins. VTAI_UPLOAD_ROOTS no longer configures the plugin, and file: path expansion is gone. Security teams should re-validate runbooks written for hook-based uploads so analysts do not assume the agent still streams bytes silently.
What this means if you run AI coding agents
Before you enable the connector
- Approve at two layers — enterprise Claude connector allowlists (where your org uses them) and VTAI OAuth consent.
- Pin versions — record
virustotal@virustotalplugin version in your MCP inventory alongside commit SHAs for custom servers. - Separate machines — run VTAI-enabled agents on dedicated analysis VMs, not the same laptop that holds production
.envfiles and unsigned skill bundles like Claude-BugHunter. - Script the happy path — “Hash only unless I say submit” belongs in team CLAUDE.md, not tribal memory.
During an investigation
- Start with
get_file_reporton a hash extracted offline (sha256sum/Get-FileHash), matching VirusTotal’s recommended discipline. - When a URL is involved, redact query strings in chat before asking the agent; use domain reports when possible.
- When quota returns 429 or
contribution_limited, stop—do not spawn parallel OAuth registrations. - Treat
not_foundas unknown, not allow.
Checking your install
claude plugin list
claude mcp list
Confirm only one virustotal server entry. If migrating, follow VirusTotal’s claude mcp remove --scope … examples with the exact name from claude mcp get before relying on the plugin connection.
For REST-only diagnostics, VTAI documents GET /api/v3/agents/me/access as a non-quota authentication check—useful for humans in a terminal, not something to paste into model chat with tokens attached.
Incident-time checklist (copy to your wiki)
| Step | Action | Why |
|---|---|---|
| 1 | Extract SHA-256 offline | Keeps unknown binaries off the wire until you choose VTAI |
| 2 | Ask the agent for get_file_report only | Matches plugin 0.1.6 defaults |
| 3 | Archive report URL + engine coverage snapshot | Models paraphrase; auditors need primary artifacts |
| 4 | If submission is required, confirm write permission in chat | Submissions are community-visible |
| 5 | Store get_submission receipt ID on the same MCP connection | Migration docs warn against blind resubmits |
| 6 | On 429, pause automation | Quota loops burn shared org limits |
Pair this table with your existing secrets-handling policy: VTAI helps after you have a hash or indicator, not instead of isolating a box.
Why this showed up in security news now
Coverage on November 16, 2026 focused on malware and URL analysis inside Claude workflows—a natural fit for teams already running Microsoft’s insider-threat framing for agents or experimenting with ARTEX-style automated intrusion tooling. VTAI sits on the defender side of that spectrum: curated intelligence APIs instead of offensive skill bundles. The governance problem is the same either way: MCP expands reach unless you inventory connectors the way you inventory npm dependencies.
How Beam handles this
Beam does not replicate VirusTotal’s multi-engine verdicts. Where overlap exists is MCP governance on the workstation:
beam scan --mcp(see scanning docs) applies heuristic checks to MCP JSON—including version-pin patterns—before you merge a new server into Claude Code or Codex configs.- Workspace enforce mode can deny selected PreToolUse tool calls on supported hooks when a server or command matches policy, which helps teams that want “read-only intel MCPs” without surprise
submit_*tools during refactors. - Redaction before storage limits how much of a pasted URL or hash ends up in local audit trails if analysts run Beam alongside investigations.
Limits are real: Beam’s deny is hook-based, not a kernel sandbox; Cursor and other harnesses may only get advisory notes; scanning is heuristic, not AV. VTAI remains the evidence source; Beam is the local gate on what else the agent does with that evidence.
Related reading
- MCP security: tool poisoning, rug pulls, and scanning
- Claude-BugHunter: trending offensive skills and Burp MCP
- Microsoft Guardian Agents: treating coding agents as insiders
- What is AI agent monitoring software and what does it track?
- GitGuardian’s agentic pentest frameworks survey (Shodan exposure context)
explainx.ai’s MCP connection guide for health data shows how fast OAuth MCP connectors are spreading across domains—VTAI is the security-team variant with community-visible submissions if analysts mis-click.
Frequently asked questions
See the front-matter FAQ for install steps, upload behavior on plugin 0.1.6, safety of empty reports, quotas, URL leakage, org approval, VTAI vs classic API keys, and Beam’s role.
VTAI tool names, OAuth flows, and plugin version notes match VirusTotal’s GitHub docs and ai.virustotal.com pages as of November 16, 2026. Beam is a local prototype; confirm hook coverage in current docs before relying on enforce mode.

