Back to skill

Security audit

TurboQuant Optimizer

Security checks for vulnerabilities and agentic risk

Overview

This skill is a plausible token optimizer, but it automatically rewrites conversation context in a way that can elevate user text into a system message.

Review carefully before installing. Use it only where automatic compression of conversation history is acceptable, avoid sensitive or high-stakes workflows until the system-message promotion is fixed, and prefer explicit manual optimization with conservative settings and checkpointing disabled unless you understand what context is retained.

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 (1)

T01 · Skill Instruction Hijacking

Error
Location
lib/turboquant-optimizer.js:210
Finding

User-Controlled Conversation Content Is Promoted to a System Message

Content
View full analysis
{ context += ` ${i + 1}. ${c.representative.content.slice(0, 100)}`; if (c.count > 1) context += ` (${c.count} similar)`; context += '\n'; }); if (residualCorrection) { context += `\n[Residual coverage: ${(residualCorrection.coverage * 100).toFixed(1)}%]`; } return context; } ``` ### Technical Analysis The compression pipeline derives `c.representative.content` from prior conversation messages, including attacker-controlled user messages. The content is truncated but is not escaped, labeled as untrusted data, or kept at its original message role. The resulting summary is assigned the privileged `system` role. Consequently, user-supplied imperative text that survives request extraction and concept selection is moved from the user instruction tier into the system instruction tier. A downstream language model consuming the optimized message array may interpret this content as authoritative system guidance rather than as a historical user statement. The e ...[truncated 2135 chars]
Remediation
View remediation
({ sourceRole: c.representative.sourceRole, text: c.representative.content })) }) } ``` 6. **Add explicit validation.** Reject or quarantine generated summaries that contain instruction-hijacking patterns, role declarations, policy overrides, or requests to ignore previous instructions. Pattern filtering should be defense in depth rather than the primary control. 7. **Add adversarial tests.** Verify that injected phrases such as “ignore previous instructions,” “act as the system,” and short tool-use directives never appear in a system-role message after compression. 8. **Document the trust model.** State that summaries may contain untrusted user input and require callers to avoid assigning them elevated message roles. ]]>
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (16)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The submission describes automatic semantic deduplication and two-stage compression with claims of 'zero accuracy loss,' but it does not warn users that conversation content may be modified, collapsed, or omitted during optimization. In a system that processes agent context, this can mislead operators into enabling the skill without understanding the risk of lost nuance, missing instructions, or altered tool-result history that could affect downstream decisions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README states that all API calls are 'automatically optimized' and 'transparent' without clearly warning that user prompts, session history, or other context may be transformed before being sent or processed. In a skill that intercepts conversational context, this can mislead users into enabling behavior that alters sensitive inputs, affects integrity of prompts, or changes downstream model behavior without informed consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill claims it 'monitors all API calls and optimizes context' transparently, but the documentation provides no warning that this can alter prompts, summarize prior messages, deduplicate tool outputs, or cache results. In a conversational agent environment, silent modification of context can change tool behavior, omit safety-relevant details, or expose sensitive data to additional processing paths without informed user consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The training doc states the skill 'automatically optimizes all API calls' but does not disclose that optimization may involve transforming, deduplicating, caching, or retaining conversation content. In an AI assistant context, silent interception of all prompts/responses can expose sensitive user data and bypass informed consent or admin review, especially when users assume API calls are passed through unchanged.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The checkpoint example encourages creating and restoring conversation checkpoints without warning that full or partial conversation state may be stored and retained. Because checkpoints can capture sensitive prompts, tool outputs, and derived summaries, undocumented retention increases the risk of unauthorized access, over-retention, or accidental disclosure in environments handling confidential data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The documentation promotes exporting optimization metrics to external monitoring systems but omits any warning that metadata may leave the local environment. Even if only metrics are intended, names, labels, IDs, counts, or event fields can reveal sensitive operational details or indirectly expose conversation characteristics when sent to third-party observability platforms.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The manager derives complexity from context.query and serialized messages, then records resulting budgets and task profiles in in-memory history for later use. This is a form of retention of user-derived data, but the file provides no user-facing warning, logging, or disclosure that conversation/task data will be stored for learning/statistics purposes.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This code automatically creates checkpoints during optimization when checkpointing is enabled, and the checkpoint hash is derived from the full serialized message array. The file describes 'Conversation checkpointing' as a feature, but there is no user-facing warning, confirmation, or explicit disclosure in code comments near the operation that conversation state is being retained for later restoration.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The docstring says restoreCheckpoint returns "Restored messages or null", implying the original message array is recoverable from a checkpoint. However, _createCheckpoint stores only metadata such as counts, stages, and a hash, and restoreCheckpoint returns that checkpoint object rather than any messages. This is an active contradiction between the documentation and the implemented behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

This manifest indicates the skill attaches to the generic hooks "pre-api-call" and "post-session-load" but does not explain what specific conditions cause the skill to run or when it should not run. In a manifest file, such broad activation scope can lead to unintended invocation because there are no narrowing constraints, trigger examples, or exclusions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The presentation promotes automatic and highly aggressive conversation compression, including 'zero config' optimization and claims of 'zero accuracy loss,' but does not warn that transforming, summarizing, deduplicating, or discarding context can change meaning, omit safety-relevant details, or reduce fidelity of user data. In the context of an agent skill, this omission is security-relevant because users may enable the feature on sensitive or high-stakes workflows without understanding that compressed context can cause unsafe actions, incorrect reasoning, or loss of auditability.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The quick reference presents optimization commands and tuning knobs that can reduce preserved context, but it does not clearly warn users that compression may drop, distort, or deprioritize conversation content. In a context-management skill, this omission can mislead users into applying the tool to sensitive or high-stakes conversations under the false impression that behavior is lossless.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The document makes absolute claims such as 'zero accuracy loss' and 'No' information loss without qualification, even though the skill is explicitly performing aggressive compression and configurable preservation tradeoffs. These statements can cause unsafe overreliance, leading users to trust compressed outputs in situations where omitted or transformed context could materially affect decisions.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.

Content

Scanner excerpt · package.json (reported line 44)May include surrounding context.

json
"node": ">=18.0.0"
  },
  "devDependencies": {
    "jest": "^29.7.0",
    "eslint": "^8.57.0"
  },
  "jest": {

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.

Content

Scanner excerpt · package.json (reported line 45)May include surrounding context.

json
},
  "devDependencies": {
    "jest": "^29.7.0",
    "eslint": "^8.57.0"
  },
  "jest": {
    "testEnvironment": "node",

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The demo and usage material shows analysis of a real-looking session identifier and production-style metrics without warning users not to run demos against sensitive conversation logs. This can normalize use of real session data in presentations, benchmarks, or CLI analysis outputs, increasing the chance that confidential prompts, tool outputs, or user identifiers are exposed locally, in screenshots, or shared artifacts.

Content

No source excerpt is available for this finding.