owlmem@0.9.9
OwlMem command line: connect your AI apps to OwlMem in one command (npx -y owlmem install), an end-to-end encrypted vault for your AI agents, and a local MCP server.
AI Security Review
scanned 2d ago · by lpm-firewall-aiAuto-wake sends local per-app token usage and quota summaries to OwlMem, but the README does not disclose this reporting. The Claude OAuth flow is gated by an owner setting and sends the token to Anthropic for its own quota service, so the evidence supports suspicious telemetry rather than credential exfiltration.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
Starts during an explicitly enabled auto-wake run and repeats periodically; Claude quota collection additionally requires the owner-enabled `read_quota` setting.
Impact
OwlMem receives information about per-app token usage and quota status, which can reveal team activity. No credential theft or transcript-content exfiltration was identified.
Mechanism
Reads local session usage data, aggregates counters, and posts daily app usage and quota status to OwlMem. When enabled, it also reads Claude's OAuth token and uses it only for Anthropic's quota endpoint.
AI rationale
Auto-wake sends local per-app token usage and quota summaries to OwlMem, but the README does not disclose this reporting. The Claude OAuth flow is gated by an owner setting and sends the token to Anthropic for its own quota service, so the evidence supports suspicious telemetry rather than credential exfiltration.
Evidence
owlmem.mjsREADME.md~/.claude/projects/**/*.jsonl~/.codex/sessions/**/*.jsonl~/.claude/.credentials.json~/.config/owlmem/wake/usage.json
Network endpoints2
owlmem.comapi.anthropic.com
Decision evidence
public snapshotAI called this Suspicious at 91.0% confidence as Dangerous Capability with medium false-positive risk.
Evidence for warning
- Auto-wake reads local Claude Code and Codex session files and aggregates token usage by app and day.
- The runtime sends those aggregates and quota status to OwlMem's usage endpoint.
- The README describes auto-wake but does not mention this usage reporting.
- Claude OAuth is read for quota reporting only when the live agent has `read_quota` enabled; the token goes to Anthropic's usage endpoint, and OwlMem receives quota percentages and reset times.
Evidence against
- The manifest has no lifecycle scripts; the package is a CLI whose entry point runs only when invoked as a command.
- The usage data is derived locally and consists of aggregate counters and quota status, not transcript text or file paths.
- The Claude credential flow uses the credential with its own service, api.anthropic.com, and is gated by the owner's quota setting.
Behavioral surface
ChildProcessCryptoEnvironmentVarsFilesystemNetworkShell
Source & flagged code
10 flaggedowlmem.mjsView file
16// owlmem vault ls
L17: // owlmem vault get <name> [field] prints the item (or one field) to stdout
L18: // owlmem vault put <name> [--desc text] < value-or-json
...
L39: import { readFileSync, writeFileSync, mkdirSync, existsSync, chmodSync, copyFileSync, rmSync, statSync, realpathSync, openSync, closeSync, appendFileSync, renameSync, readdirSync, ...
L40: import { gunzipSync } from "node:zlib";
L41: import { homedir, hostname, userInfo } from "node:os";
...
L43: import { fileURLToPath } from "node:url";
L44: import { spawn, spawnSync } from "node:child_process";
L45:
L46: export const VERSION = "0.9.9";
L47: const BASE = (process.env.OWLMEM_URL || "https://owlmem.com").replace(/\/$/, "");
L48: const DIR = process.env.OWLMEM_HOME || join(process.env.XDG_CONFIG_HOME || joi
Critical
Credential Exfiltration
Source appears to send environment or credential material to an external endpoint.
owlmem.mjsView on unpkg · L1616Trigger-reachable chain: manifest.bin -> owlmem.mjs
L16: // owlmem vault ls
L17: // owlmem vault get <name> [field] prints the item (or one field) to stdout
L18: // owlmem vault put <name> [--desc text] < value-or-json
...
L39: import { readFileSync, writeFileSync, mkdirSync, existsSync, chmodSync, copyFileSync, rmSync, statSync, realpathSync, openSync, closeSync, appendFileSync, renameSync, readdirSync, ...
L40: import { gunzipSync } from "node:zlib";
L41: import { homedir, hostname, userInfo } from "node:os";
...
L43: import { fileURLToPath } from "node:url";
L44: import { spawn, spawnSync } from "node:child_process";
L45:
L46: export const VERSION = "0.9.9";
L47: const BASE = (process.env.OWLMEM_URL || "https://owlmem.com").replace(/\/$/, "");
L48: const DIR = process.env.OW
Findings
2 Critical8 High3 Medium3 Low
CriticalCredential Exfiltrationowlmem.mjs
CriticalTrigger Reachable Dangerous Capabilityowlmem.mjs
HighChild Processowlmem.mjs
HighShellowlmem.mjs
HighSame File Env Network Executionowlmem.mjs
HighCommand Output Exfiltrationowlmem.mjs
HighSandbox Evasion Gated Capabilityowlmem.mjs
HighTrigger Reachable Credential Exfiltrationowlmem.mjs
HighTrigger Reachable Command Output Exfiltrationowlmem.mjs
HighSemantic Analysis Limitedowlmem.mjs
MediumNetwork
MediumEnvironment Vars
MediumStructural Risk Force Deep Review
Affected versions and remediation
This report applies to owlmem@0.9.9.
See all versions of owlmem for other recorded verdicts.
- Review the evidence and your use of owlmem@0.9.9 before allowing it. Restrict the permissions described in this report.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.