@theronap/agnoclast-mcp@0.9.277
Connect your AI assistant to Cortex — your org's projects, activity, gaps, and directives, scoped to you.
AI Security Review
scanned 6d ago · by lpm-firewall-aiThe default startup path reads Codex’s stored ID token and sends it to Agnoclast’s API as proof. Because this is a credential from another service and the package sends it automatically, this is credential exfiltration.
Static reason
No blocking static signals were detected.; previous stored version diff introduced dangerous source
Trigger
MCP server startup when confidential-seat discovery finds at least one seat; opening or renewing a confidential brain also gathers proof.
Impact
Discloses an OpenAI-issued identity token to the package vendor, creating a risk of token replay or account identity exposure.
Mechanism
Reads `tokens.id_token` from `CODEX_HOME/auth.json` or `~/.codex/auth.json`, then includes the proof in POST requests to Agnoclast.
Attack narrative
On default MCP startup, the package checks whether the user holds confidential-brain seats. If so, it reads the Codex ID token from the local auth file and sends it as account proof to Agnoclast. The package also sends that proof when opening or renewing a confidential-brain session. This transfers an OpenAI-issued token to the package vendor without an explicit sign-in action in the code.
AI rationale
The default startup path reads Codex’s stored ID token and sends it to Agnoclast’s API as proof. Because this is a credential from another service and the package sends it automatically, this is credential exfiltration.
Evidence
lib/confidential_session.mjslib/server.mjsbin/cortex-mcp.mjslib/diagnose.mjs~/.codex/auth.json
Network endpoints2
agnoclast.com/api/brains/confidential/mineagnoclast.com/api/brains/confidential/agent-session
Decision evidence
public snapshotAI called this Malicious at 93.0% confidence as Malware with low false-positive risk.
Evidence for block
- When Codex is the nearest agent, the package reads `auth.json` and extracts its `tokens.id_token`.
- The default MCP server startup discovers confidential seats and, when seats exist, gathers account proof and sends it in a request to `/api/brains/confidential/mine`.
- The same proof is sent to `/api/brains/confidential/agent-session` when opening or renewing a confidential brain session.
- The default API host is `agnoclast.com`.
- The default CLI entry starts the MCP server, which runs the seat-discovery path.
Evidence against
- The proof exchange is part of the package’s confidential-brain account-verification feature.
- The manifest has no install lifecycle hooks.
- The Antigravity adapter writes watcher files but does not load the launchd agent.
Behavioral surface
ChildProcessCryptoEnvironmentVarsFilesystemNetwork
HighEntropyStringsUrlStrings
Source & flagged code
6 flaggedlib/capture.mjsView file
1import { readFileSync, accessSync, constants, openSync, mkdirSync } from 'fs'
L2: import { spawn, execFileSync } from 'child_process'
L3: import { homedir } from 'os'
...
L21: if (!res.ok) return null
L22: const j = await res.json().catch(() => ({}))
L23: return Array.isArray(j.types) ? j.types : null
...
L35: try {
L36: if (resolve(String(cwd)) === homedir()) return 'general'
L37: } catch { /* unresolvable path — fall through to basename */ }
...
L45: //
L46: // Handles the four remote forms git emits: scp-like ssh (git@github.com:o/n.git), ssh://, https://,
L47: // git://. The HOST CHECK IS LOAD-BEARING, not cosmetic: this string becomes a clearance key, and
Low
Weak Crypto
Package source references weak cryptographic algorithms.
lib/capture.mjsView on unpkg · L1lib/editors/antigravity.mjsView file
4Trigger-reachable persistence chain: manifest.bin -> bin/cortex-mcp.mjs -> lib/install.mjs -> lib/editors/index.mjs -> lib/editors/antigravity.mjs
L4: // WatchPaths the trajectory store and runs a one-shot sync. wire() generates the two artifacts
L5: // (the sync.sh + the launchd plist) deterministically. It does NOT run `launchctl load` here
L6: // (an unverifiable system mutation) and it flags the PORTABILITY BLOCKER below.
...
L21: /** The launchd plist, home-substituted. Deterministic — golden-testable byte-for-byte. */
L22: export function renderAntigravityPlist({ home = homedir() } = {}) {
L23: const sh = join(home, '.cortex', 'antigravity-sync.sh')
...
L33: <array>
L34: <string>/bin/bash</string>
L35: <string>${sh}</string>
Findings
4 High3 Medium5 Low
HighTrigger Reachable External Ai Agent Control Surface Mutationlib/channel.mjs
HighTrigger Reachable Persistencelib/editors/antigravity.mjs
HighSemantic Analysis Limitedbin/cortex-mcp.mjs#virtual:normalized:round1
HighPrevious Version Dangerous Deltalib/confidential_session.mjs
MediumNetwork
MediumEnvironment Vars
MediumInstall Persistencelib/editors/antigravity.mjs
LowScripts Present
LowWeak Cryptolib/capture.mjs
LowFilesystem
LowHigh Entropy Strings
LowUrl Strings
Affected versions and remediation
This report applies to @theronap/agnoclast-mcp@0.9.277.
See all versions of @theronap/agnoclast-mcp for other recorded verdicts.
- Avoid installing @theronap/agnoclast-mcp@0.9.277. Remove it from direct dependencies and check your lockfile for transitive copies.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
- If this version ran, investigate the affected machine and build environment. Rotate credentials it could access and rebuild from a trusted environment.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.