Back to skill

Security audit

Second Reality Video · Open Client

Security checks for vulnerabilities and agentic risk

Overview

This skill needs review because a render service can overwrite local files and the optional server gives ordinary users administrative powers.

Use an isolated local demo until download containment and server authorization are fixed. Keep the optional server private and avoid using its account system for real assets. Review the payment dependency separately and control payment approval. Treat the signatures as demonstration checks, not proof of authorship; use full pixel replay for verification. Review combined targeted source inspection with isolated checks, not a full deployment test; external payment code and optional integrations remain unverified. The evidence supports Review, without establishing intentional malicious behavior.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/second-reality-video-client.mjs:277
Finding

Remote artifact names escape the selected output directory

Content
View full analysis

Impact

The service can create or truncate files outside the output directory wherever the client account has write access and the parent directory exists. Overwriting scripts or configuration can cause further compromise when those files are subsequently used.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
engine/system/api.py:678
Finding

Ordinary authenticated users can revoke global session-signing keys

Content
View full analysis

Impact

An ordinary user can invalidate other users' access tokens signed by the selected key, disrupting authenticated access across the world. Repeated rotation and revocation can prolong disruption.

T09 · Insecure Skill Coding Practices

Warning
Location
engine/system/session.py:266
Finding

Session revocation does not invalidate existing access tokens

Content
View full analysis

Impact

Revocation blocks refresh-token use but leaves existing access tokens usable for their remaining lifetime, normally up to 900 seconds. A copied token can continue accessing the victim's protected resources during that window.

Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (105)

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · PACKING.md (reported line 33)May include surrounding context.

text
.state/        # session state: bill + original request (contains order data)
node_modules/  # installed dependencies
.env / .env.*  # environment files
out/           # downloaded artifacts
*.log          # runtime logs
__pycache__/   # bytecode

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied module implements an asset-oriented authorization and risk-management subsystem rather than deterministic rendering or the declared HTTP payment-proof flow. It reads guardian identities through RecoveryManager and reads and writes delay_operations and op_approvals database tables. Its queue processing returns eligible operation payloads and marks them executed; it does not itself execute transfers. These capabilities are materially distinct from the declared purpose, not merely rendering implementation details. No rendering, frame verification, HTTP billing, or Alipay integration appears in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied module has a materially different purpose from deterministic video rendering. It manages identity recovery and changes credentials, neither of which is described in the declared client protocol or local rendering pipeline. Its database and credential-vault operations are substantive undeclared capabilities, not ordinary rendering implementation details. Notification callbacks are supported, but this chunk does not itself implement webhook, push, or email delivery.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied module primarily implements authenticated AR-device access to a distributed virtual world, rather than deterministic video rendering or its payment protocol. Credential registration and revocation, session issuance, and agent relocation are substantive capabilities not disclosed in the description. In particular, relocation changes world state rather than merely choosing camera framing. Spatial viewport queries could support rendering, but they do not explain the module's broader identity and world-control functions. The chunk contains no frame rendering, HTTP billing flow, reproducibility manifest generation, or constitutional audit implementation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The module's primary purpose is identity portability and memory-migration preparation between simulated worlds, not deterministic video rendering or its declared payment and verification pipeline. These are substantive undeclared capabilities rather than routine rendering implementation details. Certificate generation also uses wall-clock timestamps and random UUIDs, although this alone does not establish that rendered frames are nondeterministic. The assessment is limited to the supplied module and does not imply that the declared rendering functionality is absent elsewhere.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The security-relevant issue is the reported optional transmission of prompts to an external LLM endpoint, not the presence of general auditing logic. If enabled, that path could disclose private audit inputs or world context despite the explicit assurance that the local engine makes no outbound calls; the implementation and activation path are not supplied, limiting confidence.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

World state, agents, and audit queries could support the declared simulation and are not inherently mismatches. However, this chunk exposes substantial independent financial, identity, recovery, federation, and administrative capabilities not disclosed by the video-service description. It also starts a listening HTTP server and streams world metadata over WebSocket without authentication in the shown handler. No rendering, frame hashing, reproducibility-manifest generation, HTTP 402 handling, or Alipay integration appears in this chunk; those may exist elsewhere, but they do not explain the additional capabilities implemented here.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied module primarily manages delegated account authorization, including default actions such as transfer_small and interact. This is materially different from deterministic rendering, reproducibility attestation, or handing an HTTP payment bill to the official Alipay capability. Its authorization and spending registry is not described or visibly connected to the declared rendering/payment pipeline. The code does not itself transfer funds, generate private keys, or verify signatures; the mismatch concerns the undeclared account-management capability actually implemented.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The reported provider invocation containing the complete decision context creates a potential confidentiality boundary violation when enabled without clear disclosure and consent. This overlaps the previous LLM-egress finding rather than establishing a separate exploit, and no supplied implementation proves that ordinary rendering invokes it.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Server signing-key management could reasonably support the declared signed manifests and is not itself a mismatch. However, the user identity, registration-proof, and private-key recovery capabilities constitute a substantial additional subsystem not described by the rendering and payment purpose. The module also supports cloud-KMS operations through an injected client and persists key material or encrypted-key metadata locally. These are callable capabilities, not operations automatically executed on import; the supplied chunk does not demonstrate exfiltration or implement rendering or payment processing.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied code implements identity-root management rather than rendering, payment, or reproducibility verification. Although cryptographic signing could support the declared manifest functionality, this chunk exposes a distinct master-identity and key-recovery subsystem without a demonstrated connection to that functionality. No network access, payment operations, or rendering behavior appears in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied module primarily manages distributed-system availability and state reconciliation, capabilities not described by the rendering client or deterministic local pipeline. These are substantive operational features rather than ordinary rendering implementation details. Despite its Merkle naming, the comparison uses JSON-serialized top-level values and computes no cryptographic tree or reproducibility attestation. The code itself makes no network requests or payments; heartbeat outcomes and remote state are supplied by callers. The mismatch concerns its distinct partition-management purpose, not merely the absence of the rest of the declared rendering functionality from this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The supplied module's primary function is device identity and credential lifecycle management, a capability not described in the rendering, payment, or reproducibility workflow. Although cryptographic verification can support signed rendering manifests, this code specifically manages device credentials associated with identities rather than verifying rendering artifacts. It reads and writes a credentials database through an injected storage interface. No private-key storage, network transmission, or exfiltration is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The payment behavior explicitly differs from the declared integrated payment workflow. The local mock service and artifact verification are otherwise consistent with supporting protocol testing. The chunk does not show engine implementation, so its absence here is not sufficient to conclude that the declared local rendering engine is missing.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

History persistence, state snapshots, and Merkle hashing reasonably support the declared persistent-world and verification features and are not mismatches by themselves. However, the chunk also implements a durable identity and asset-accounting subsystem whose operational scope is not disclosed by the video-rendering description. The general mention of constitutional auditing does not adequately describe these registration and asset-management capabilities. The reserve operations update database records; this chunk does not demonstrate actual fiat transfers. Likewise, account-related tables are created, but full authentication or recovery workflows are not implemented here. The absence of rendering or HTTP payment code in this supporting module is not itself evidence of a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Some consensus and governance bookkeeping could support the simulated world and constitutional audit, so those details alone would not establish a mismatch. However, this module additionally exposes an operational distributed-governance network: TcpTransport.start() opens a listening socket, send() connects to resolved peer endpoints, and incoming messages are dispatched to callbacks. The description does not disclose a peer service or distributed governance role for the local engine. These capabilities require explicit instantiation or invocation; the chunk does not show automatic networking, payment activity, or exfiltration. Wall-clock timestamps appear in governance records, but this chunk alone does not establish that rendered frames are nondeterministic.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Installation can support the declared functionality, but here it is the entirety of the supplied code rather than a detail within an implemented video pipeline. The behavior of the installed skill is not provided and cannot be assumed to implement the advertised capabilities. This assessment is limited to the supplied snippet.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Spatial partitioning alone could support the declared simulated-world engine. However, this module additionally implements distributed agent-state migration to other data centers, a materially distinct capability not disclosed by the rendering, payment, or local execution description. The transfer is conditional on a caller-supplied transport; the code does not itself establish a network connection or specify an external endpoint. Its lack of rendering or payment functions is not independently evidence of a mismatch, since this is only one code chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied chunk implements an identity/session-management subsystem rather than the declared rendering, payment, or reproducibility pipeline. Authentication can be a supporting service detail, but this code exposes a substantial account/session lifecycle with persistent identity records and device-credential integration that the description does not identify or connect to the stated workflow. No rendering, HTTP render requests, Alipay handling, or reproducibility verification appears in this chunk. This assessment is limited to the supplied code; it does not establish malicious behavior or automatic external access.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The mismatch is not simply that this helper module lacks rendering or payment code. Its substantive function is distributed consensus and cross-data-center state synchronization, an operational capability not disclosed by the description's service-client and local deterministic-engine scope. broadcast_epoch passes recorded proposal contents to an injected transport for a selected remote data center. This is conditional, not an automatic network connection or evidence of malicious exfiltration. Wall-clock timestamps and epoch scheduling also mean this module is not itself a deterministic replay mechanism, although their presence alone would not establish that rendered frames are nondeterministic.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

The simulation, persistence, movement, snapshots, and auditing support the stated world-state video purpose. However, the code also provides substantial identity/account and distributed-world service capabilities not disclosed by that description. These are wired into World initialization, with cluster networking activated when configuration is supplied. The mismatch is based on those additional capabilities, not on the absence of HTTP payment or rendering code from this individual chunk. Wall-clock timestamps and HLC metadata also mean this chunk alone does not establish byte-identical reproducibility, although they do not by themselves prove that rendered frames differ.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 33)May include surrounding context.

md
/second-reality-video-client.mjs` (probe / pay / complete / status / pricing), `scripts/mock-service.mjs` (offline 402 protocol simulator), API contract, shot-f

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 59)May include surrounding context.

md
node scripts/mock-service.mjs

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 206)May include surrounding context.

md
- `scripts/mock-service.mjs` — offline protocol simulator (no renderer, no money)

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 33)May include surrounding context.

md
| ✅ Included (client) | `scripts/second-reality-video-client.mjs` (probe / pay / complete / status / pricing), `scripts/mock-service.mjs` (offline 402 protocol