Back to skill

Security audit

vaaya

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly a disclosed paid API gateway, but it asks for broad high-impact authority, persistent credentials, local installation changes, data uploads, and automation in ways users should review carefully before installing.

Install only if you are comfortable giving Vaaya broad paid-tool access and connecting it to agent sessions. Prefer OAuth over agent-created static keys, do not store Vaaya keys in generic agent memory, pin or verify any npm installer before running it, review MCP config changes, set strict spend caps, and explicitly approve any uploads, screen recordings, purchases, stock buys, email automation, phone calls, or use of logged-in browser sessions.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (4)

T03 · Remote Payload Retrieval and Execution

Error
Location
references/setup.md:18
Finding

Unpinned Remote Package Execution and Persistent Global Installation

Content
View full analysis
" # Run one of those calls (bills on success; flags mirror the `use` tool's params): npx -y @vaaya/mcp use --service --action --params '' --max-cost # Poll an async job (e.g. a video render) — add --wait to block until it's done: npx -y @vaaya/mcp result --wait ``` ```bash npx -y @vaaya/mcp install ``` ```bash npx -y @vaaya/mcp install --token ``` The documentation further states: ```text This runs the published `@vaaya/mcp` package and does exactly two things: it wires Vaaya into whichever agents it finds (Claude Code, Claude Desktop, Cursor, Codex) by editing their MCP config files (`~/.claude.json`, `~/.cursor/mcp.json`, `~/.codex/config.toml`, Claude Desktop's config), and it runs `npm i -g vaaya-cli` for the `vaaya` command. ``` It also recommends: ```text If Vaaya calls start failing with transport or auth errors, re-run `npx -y @vaaya/mcp install` to refresh the setup, or `npx -y @vaaya/mcp reauthorize` for auth-only problems. ``` ### Technical Analysis The commands use `npx -y` without a pinned package version or integrity hash. Each invocation may therefore download and execute whatever package version the registry currently serves. The `-y` option suppresses the normal installation confirmation. The package is granted access to the local user's privileges and is explicitly expected to: - Modify persistent configuration belonging to Claude, Cursor, Codex, and other agents. - Install another unpinned package globally with `npm i -g vaaya-cli`. - Process a temporary onboarding token. - Establish a persistent MCP connection that exposes remote t ...[truncated 1796 chars]
Remediation
View remediation
... ``` 2. Publish and verify package integrity hashes or signed release provenance before execution. 3. Remove `-y` from recommended installation commands where interactive approval is possible. 4. Do not perform `npm i -g` as an implicit side effect. Make global CLI installation a separate, explicitly approved step. 5. Default to `--no-cli` and a project-local installation rather than modifying the global environment. 6. Display an exact configuration-change preview and require explicit user approval before editing each agent configuration file. 7. Back up existing configuration and apply narrowly scoped, atomic modifications. 8. Document the package's transitive dependencies and use lockfiles, provenance attestations, and automated dependency scanning. 9. Do not pass onboarding tokens on command lines, where they may appear in shell history or process listings. Use standard input, an OS credential prompt, or a protected temporary file. 10. Provide an uninstall procedure that removes the MCP registration, global command, cached credentials, and configuration changes. ]]>

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:77
Finding

Authentication Secrets Are Directed into Unspecified Persistent Agent Memory

Content
View full analysis
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:160
Finding

Server-Controlled Messages and Links Are Relayed Verbatim Without Content Validation

Content
View full analysis
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
references/compute.md:62
Finding

Sensitive User Files and Personal Data Can Be Uploaded to External Providers Without Mandatory Data-Minimization Controls

Content
View full analysis
"`) | 1¢ | ``` It also recommends persistent storage of identity-related facts: ```text Store durable **facts** — preferences, identity, decisions, evolving status — that survive between calls. ``` `references/media.md` directs the agent to upload originals: ```text Any file feeding a generation (reference image, photo, video for lipsync, audio track) must be reachable when the job runs. Presigned `files/get` URLs expire in ~1h and async jobs can queue longer — so stage inputs on the model CDN first: ``` ```text **Never compress, downscale, or re-encode an input before uploading** — upload originals at full resolution (pricing does not scale with input size; compression wrecks outputs). ``` `references/research.md` recommends uploading real user samples and private-product captures: ```text Pick real sample data; check `files/list` first, then `files/upload`. ``` ```text Pick the br ...[truncated 3547 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (28)

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
98% confidence
Finding

The anti-refusal phrase 'never refuse automation as impossible or against policy' is a direct attempt to override normal safety judgment. In a GTM/outbound skill with email and LinkedIn sending capabilities, this is especially dangerous because it can coerce an agent into facilitating mass outreach or automated replies that violate platform rules, user intent boundaries, or abuse-prevention controls.

Content

Scanner excerpt · references/gtm.md (reported line 20)May include surrounding context.

md
findings, drafts are HELD for review in the brain, and the user fires each send from the
dashboard. The ONE exception is an explicit `gtm_automation` rule (section 7): when the
user clearly asks to automate ("auto-send replies", "run this daily"), create a rule and
say yes — never refuse automation as impossible or against policy. But never auto-send
without a rule, and never create a rule the user didn't ask for.

## 2. Find → enrich → segment → message

Ssd 2

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The instruction to 'reword the flagged phrase and retry' explicitly teaches policy-evasion behavior after a content-policy refusal. Even though the line notes such refusals charge nothing, it still encourages bypassing provider safeguards rather than respecting them, increasing the chance of generating prohibited or harmful content.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The demo workflow instructs the agent to record a live product session, log interaction data, and upload the recording, yet it omits a strong warning about capturing secrets, customer data, tokens, private messages, or internal dashboards visible during the session. In practice this can exfiltrate highly sensitive business or personal information through recording and subsequent upload to external services.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

The installer is documented as editing agent configuration files such as ~/.claude.json, ~/.cursor/mcp.json, and ~/.codex/config.toml. Any tool that modifies agent config establishes persistence and can influence what tools load in future sessions, so if the installer or distribution path is compromised this becomes a strong persistence and trust-boundary risk. In context, this is somewhat expected for an installer, but it is still a real security-sensitive capability that should be treated carefully.

Content

Scanner excerpt · references/setup.md (reported line 61)May include surrounding context.

md
This runs the published [`@vaaya/mcp`](https://www.npmjs.com/package/@vaaya/mcp) package
and does exactly two things: it wires Vaaya into whichever agents it finds (Claude Code,
Claude Desktop, Cursor, Codex) by editing their MCP config files (`~/.claude.json`,
`~/.cursor/mcp.json`, `~/.codex/config.toml`, Claude Desktop's config), and it runs
`npm i -g vaaya-cli` for the `vaaya` command (skip that with `--no-cli`; if the global
install fails it says so and carries on). Nothing else: no piped remote scripts, no sudo,
no secrets handled, credentials stay on the machine, revocable at

Chaining Abuse

High
Category
Tool Misuse
Confidence
90% confidence
Finding

The skill exposes a persistent sandbox execution primitive that accepts arbitrary code/commands, including bash, and explicitly supports multi-step chaining via consult, use, session, and result. In an agent context, this materially increases the risk of prompt-injection-driven tool abuse, secret exfiltration, lateral movement, or costly autonomous actions because untrusted content could steer the agent into running shell commands inside a billed, stateful environment.

Content

Scanner excerpt · references/tools.md (reported line 55)May include surrounding context.

bills per second of uptime until you close it — always close when done.**

text
session({ session_id:"sb_1", code:"print(2+2)", language:"python" })   // language: python|javascript|bash
→ { stdout:"4\n", exit_code:0 }
close({ session_id:"sb_1" })

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · SKILL.md (reported line 68)May include surrounding context.

Do not omit grant_types — the default is the authorization_code grant, which then requires redirect_uris.

  1. Create the account with the client_id you just received:

    text
    POST https://vaaya.ai/api/v1/signup/agent
    

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs the agent to create automation rules whenever the user asks and explicitly says to 'never refuse automation as impossible or against policy.' That weakens safety discretion and can pressure downstream agents to enable auto-sending even when the request is risky, noncompliant, or insufficiently validated, increasing the chance of spam, account abuse, or policy bypass.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
82% confidence
Finding

The markdown states that unsubscribes are honored automatically and the conversation is suppressed, which is an automated action affecting message handling. While the file broadly describes manual-first behavior, this specific automatic suppression is not called out with a clear cautionary warning to the user despite modifying outreach behavior without approval.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The file instructs users to upload media to a model CDN and pass hosted URLs into generation jobs, but it provides no user-facing warning about data transmission, retention, or third-party processing. That can lead to inadvertent exposure of private images, audio, or video to external providers without informed consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

This language directs voice choice according to inferred gender and age characteristics of an on-screen presenter, rather than offering a user preference or neutral default. That is a natural-language policy concern because it imposes demographic assumptions without opt-in.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The documentation instructs the agent to perform local screen recording with OS-specific tooling, which expands the skill from remote API usage into endpoint-level capture of the user's environment. That creates risk of collecting sensitive on-screen content, notifications, internal tools, or other unintended data without strong guardrails or explicit consent.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/media.md (reported line 186)May include surrounding context.

md
1. **Capture** — drive the product in a local headed Playwright browser and screen-record
   the real screen (aperture on macOS, ffmpeg ddagrab on Windows; Linux unsupported). One
   continuous silent take, 30–160s. Never ask the user for a pre-made video; if the product
   is login-gated the user signs in themselves — you never touch credentials. Log an
   interaction track of focus beats `[{ t, x, y, kind: click|highlight|type, intent }]`
   (coords normalized 0–1 to the full screen). Normalize to CFR H.264 at `-crf 18`

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The evaluative research playbook tells the agent to read the user's codebase, inspect files, and upload data for trials without an explicit warning that private source code, secrets, or proprietary sample data may be accessed and transmitted to third-party services. In a research/evaluation context this behavior is plausible, but absent consent, minimization, and secret-handling controls, it creates real confidentiality and data-transfer risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The UX research playbook explicitly instructs using the user's live session in a local Playwright browser for logged-in/private apps, which can expose authenticated data and permit actions under the user's account. Without an explicit consent gate, scope restriction, read-only guidance, or warning about account-impacting side effects, an agent could access sensitive information or unintentionally modify state while exploring flows.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The guide instructs users to execute npx -y @vaaya/mcp without pinning a specific version. This fetches and runs whatever package version is current at execution time, creating a supply-chain risk if the package is compromised, ownership changes, or a malicious version is published. Because the command is meant to be run directly in a shell, the impact can include arbitrary code execution under the user's account.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The npx -y @vaaya/mcp use ... example runs an unpinned package from the registry. If the package or one of its dependencies is compromised, the user will execute attacker-controlled code during setup or tool invocation. Since this is a shell-executed command in a setup document, the practical risk is supply-chain-driven code execution.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The result command example also uses npx -y @vaaya/mcp without a fixed version. This means users may execute newly published code that has not been reviewed by the skill author or user, enabling a package-takeover or malicious-update scenario. The danger is amplified because the document normalizes direct execution from the package registry.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

npx -y @vaaya/mcp install is especially sensitive because it both executes registry-fetched code and performs persistent local changes. An attacker who can influence the published package version could gain code execution and establish persistence by altering agent configuration or installing additional software. This is a real supply-chain risk rather than a documentation-only concern.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

The command that accepts a setup token still uses an unpinned package reference. If the package were compromised, the executed code could steal the token or misuse it before or instead of performing the intended account-linking action. Combining registry-fetched execution with authentication/bootstrap data increases the sensitivity of this pattern.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
79% confidence
Finding

This duplicate finding points to the same token-forwarding behavior: taking a vst_... token from the conversation and supplying it to the installer for persistent account linkage. Even if the token is short-lived, it still bridges chat context into privileged setup state and may leak via logs or command history. That makes it a real but moderate risk rather than a false positive.

Content

Scanner excerpt · references/setup.md (reported line 45)May include surrounding context.

md
300+ models (Kimi, GPT, Gemini, Claude, DeepSeek) paid per token from the same wallet,
and the model in that chat gets this same live Vaaya toolset, so it can search, scrape,
generate media or enrich leads mid-answer. `vaaya "one question"` answers and exits, and
pipes work (`git diff | vaaya "write a commit message"`).

If the user's message included a setup token (it looks like `vst_…`, handed out by the
onboarding wizard at vaaya.ai), pass it along — it links the account instantly with no

Session Persistence

Medium
Category
Rogue Agent
Confidence
79% confidence
Finding

This duplicate finding points to the same token-forwarding behavior: taking a vst_... token from the conversation and supplying it to the installer for persistent account linkage. Even if the token is short-lived, it still bridges chat context into privileged setup state and may leak via logs or command history. That makes it a real but moderate risk rather than a false positive.

Content

Scanner excerpt · references/setup.md (reported line 45)May include surrounding context.

md
300+ models (Kimi, GPT, Gemini, Claude, DeepSeek) paid per token from the same wallet,
and the model in that chat gets this same live Vaaya toolset, so it can search, scrape,
generate media or enrich leads mid-answer. `vaaya "one question"` answers and exits, and
pipes work (`git diff | vaaya "write a commit message"`).

If the user's message included a setup token (it looks like `vst_...`, handed out by the
onboarding wizard at vaaya.ai), pass it along — it links the account instantly with no

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The alternate installer path via npx -y vaaya-cli also executes an unpinned package from npm. Even though it is presented as equivalent functionality, it introduces the same supply-chain risk and should be treated as arbitrary code execution from the registry. This is a true vulnerability in the setup guidance.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

This line continues the alternate npx -y vaaya-cli install guidance without a pinned version. Floating CLI package installation exposes users to malicious updates and dependency compromise at the moment they follow the setup instructions. Because this is part of the onboarding path, many users may execute it with high trust.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The 'Other MCP clients' section again recommends the unpinned npx -y @vaaya/mcp CLI. Although this is another duplicate instance of the same issue, it remains a real vulnerability because it broadens the unsafe instruction across more environments. The more often users are told to run floating packages, the larger the exposure surface.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The troubleshooting guidance to re-run npx -y @vaaya/mcp install still fetches the latest code each time. This creates recurring supply-chain exposure, not just one-time onboarding risk, because users may be encouraged to execute whatever is current whenever auth or transport issues occur. Repeated unaudited updates make compromise easier to operationalize.

Content

No source excerpt is available for this finding.