Back to plugin

Security audit

Memory (SynapCores)

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this plugin suspicious because scanners report that deleted memories may remain retrievable and remote connections may expose credentials and memory data over unencrypted HTTP.

Review the plugin before installing. If you use it, check that any remote gateway uses HTTPS and consider whether its memory deletion behavior is acceptable, since deleted data may remain in graph copies under the reported conditions.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (2)

MCP01

Warning
Location
plugin/dist/index.js:2655
Finding

Scope-Based Forgetting Can Leave Retrievable Graph Copies While Reporting Successful Cleanup ## Remediation Suggestions 1. Have the authoritative scope-deletion operation return all deleted episode IDs, or maintain a durable mapping between memory records and their graph copies. Do not use limited semantic recall as the deletion inventory. 2. Coordinate memory and graph deletion transactionally where supported. Otherwise, enqueue durable, idempotent cleanup jobs with retries and reconciliation. 3. Return an incomplete or unknown cleanup status when inventory collection fails. Never infer complete cleanup solely from the absence of recorded removal failures. 4. Ensure recalled fact IDs are not incorrectly substituted for the episode IDs used as graph-node keys. 5. Test deletion involving more than 25 records, inventory failures, differing fact and episode IDs, and graph-service outages. Verify that forgotten text cannot subsequently be retrieved from graph copies.

Content
View full analysis

Vulnerability Details

File Location: plugin/dist/index.js:2655–2678
Vulnerability Type: Incomplete deletion of sensitive data
Risk Level: Medium

Vulnerable Code

js
let candidateIds = [];
try {
    const pre = await router.recallForPrompt(query, identity, 25);
    candidateIds = pre.results.map((r) => r.entry.id).filter(Boolean);
}
catch {
    // best-effort: cleanup proceeds without pre-collected ids
}
let scoped;
try {
    scoped = await router.forgetAbout(query, identity);
}
catch (err) {
    return asToolError(err);
}
if (scoped.supported) {
    // Explicitly remove the graph copies of the forgotten memories so
    // a scope-based forget leaves nothing recallable in the graph.
    let graphFailures = 0;
    for (const cid of candidateIds) {
        if (!(await removeGraphCopy(client, cid, api.logger)))
            graphFailures++;
    }
    const affected = (scoped.affected ?? {});

The subsequent response construction reports:

js
graphCleanup: graphFailures > 0 ? `failed:${graphFailures}` : "removed",

Technical Analysis

When autoLinkSimilar is enabled, the plugin creates separate graph nodes containing memory text and embeddings. The memory backend's scope-based deletion does not automatically remove these plugin-created copies.

The handler attempts to inventory graph copies using semantic recall limited to 25 results, then performs a separate scope-based deletion. There is no guarantee that the recalled IDs exhaustively identify the records deleted by that operation. A scope deletion can therefore remove more records than the graph-cleanup loop processes.

The exception handler also silently discards failures during the inventory step. In that case, candidateIds remains empty, no graph cleanup is attempted, and graphFailures remains zero. The response can consequently report graph cleanup as "removed" without verifying that the graph contains no surviving copies.

This is a confirmed implementation defec ...[truncated 1761 chars]

other

Warning
Location
plugin/dist/index.js:2393
Finding

Remote Gateway Connections Permit Cleartext Transmission of Credentials and Memory Data ## Remediation Suggestions 1. Reject non-loopback HTTP destinations before constructing the SDK client or initiating network operations. 2. If insecure remote transport must remain supported, require a separately named, explicit override that clearly acknowledges credential and memory exposure. 3. Default remote gateway configurations and documentation examples to HTTPS. 4. Verify that the SDK validates certificates and hostnames, rejects HTTPS-to-HTTP redirects, and does not forward credentials to unrelated redirect destinations. 5. Use narrowly scoped gateway credentials. Rotate credentials that may already have traversed an unprotected network. 6. Add configuration tests confirming that remote HTTP is rejected by default and that sensitive values are not included in transport-error logs.

Content
View full analysis

Vulnerability Details

File Location: plugin/dist/index.js:2393–2397; related configuration at plugin/dist/config.js:224–248
Vulnerability Type: Cleartext transmission of sensitive information
Risk Level: Medium

Vulnerable Code

The configuration parser enables HTTPS only when explicitly requested:

js
const useHttps = synapcores.useHttps === true;

The runtime passes that setting and the configured credential to the SDK:

js
const client = new SynapCores({
    host: cfg.synapcores.host,
    port: cfg.synapcores.port,
    useHttps: cfg.synapcores.useHttps,
    apiKey: cfg.synapcores.apiKey,
});

For remote HTTP connections, the transport guard only logs a warning:

js
if (!cfg.synapcores.useHttps && transportHost && !isLoopbackHost) {
    api.logger.warn(`memory-synapcores: connecting to remote gateway '${transportHost}' over plaintext ` +
        `HTTP — memory traffic and the API key are unencrypted on the network path. Set ` +
        `synapcores.useHttps: true for any non-loopback host.`);
}

Technical Analysis

HTTPS is opt-in rather than required for remote destinations. An operator can configure a non-loopback gateway while omitting useHttps: true; the plugin then continues with HTTP despite detecting the insecure transport.

Memory operations use the configured gateway credential and transmit stored or recalled conversation-derived information. Without TLS or independently protected transport, an observer on the network path can read that traffic. An active on-path attacker can potentially modify requests or responses.

This is a deployment-dependent transport weakness, not evidence of deliberate exfiltration. The default loopback destination does not itself establish external network exposure. SDK behavior, including certificate validation and redirects, was outside the audit scope.

Attack Path

Prerequisites: The operator configures a remote gateway over HTTP, the network path lacks equivalent ...[truncated 1398 chars]

SkillSpector was not run because this plugin release contains no bundled skills.

Endor dependency reachability

Checked Nov 15, 2026 · 4:44 PM UTC

No reachable vulnerabilities reported

Endor dependency scan completed.