cosmos-vulnerability-scanner

Cosmos Vulnerability Scanner

Purpose

Scan Cosmos SDK modules and CosmWasm contracts for vulnerabilities that cause chain halts, consensus failures, or fund loss. Spawns parallel scanning agents — each specializing in a vulnerability category — that return findings to the main skill, which then writes them as individual markdown files to an output directory.

Output directory: defaults to .bughunt_cosmos/. If the user specifies a different directory in their prompt, use that instead.

When to Use

• Auditing Cosmos SDK modules (custom x/ modules)
• Reviewing CosmWasm smart contracts
• Pre-launch security assessment of Cosmos chains
• Investigating chain halt incidents

When NOT to Use

• Pure Solidity/EVM audits without Cosmos SDK — use Solidity-specific tools
• CometBFT consensus engine internals — this covers SDK modules, not the consensus layer itself
• General Go code review with no blockchain context
• Cosmos SDK application logic that is not consensus-critical (e.g., CLI commands, REST endpoints)
• CosmWasm contract-only audits on chains without custom SDK modules — use the CosmWasm checklist items alone

Essential Principles

1. Consensus path is king — A bug only matters for chain halt/fund loss if it's on the consensus-critical execution path (BeginBlock, EndBlock, FinalizeBlock, msg_server handlers, AnteHandler). Always verify a finding is reachable from consensus before reporting it.
2. State divergence = chain halt — Any non-determinism that causes validators to compute different state roots will halt the chain. This is the highest-severity class because it affects all validators simultaneously.
3. Check the version — Cosmos SDK has breaking changes across major versions (v0.47 removed GetSigners, v0.50 added ABCI 2.0, v0.53 deprecated ValidateBasic). Always check go.mod versions before applying patterns.
4. False positives waste audit time — A map iteration in a CLI command is not a consensus bug. A panic in a query handler does not halt the chain. Verify the execution context before flagging.
5. Cross-module interactions are where bugs hide — The most severe findings (IBC reentrancy, EVM/Cosmos state desync, authz escalation) involve interactions between modules, not bugs within a single module.

Scanning Workflow

Phase 1: Discovery (synchronous)

Entry: Target codebase path provided by user. Codebase contains Go source (e.g., x/ modules, go.mod) or Rust contracts with cosmwasm_std.

Run a synchronous subagent (Agent tool) with the full contents of DISCOVERY.md as its prompt. The agent must:

1. Follow the Discovery workflow to explore the target codebase
2. Return the full CLAUDE.md content (the technical inventory and threat model) in its response
3. Return a structured summary with exactly these fields:

PLATFORM: pure-cosmos | evm | wasm        (pick one; if multiple, comma-separated)
IBC_ENABLED: true | false
SDK_VERSION: <version from go.mod>
IBC_GO_VERSION: <version from go.mod, or "n/a">
CUSTOM_MODULES: <comma-separated list of x/* modules>

After the subagent returns, you (the main skill) Write the CLAUDE.md to the target repo root. Save its path and the discovery values — these feed into Phase 2.

Exit: CLAUDE.md written by main skill. PLATFORM, IBC_ENABLED, SDK_VERSION, IBC_GO_VERSION, and CUSTOM_MODULES captured.

Phase 2: Parallel Vulnerability Scan

Spawn scanning agents in a single message for maximum parallelism. Use the Agent Prompt Template below, filling in the reference file for each agent. Subagents only need read access (Grep, Glob, Read) — they return findings in their response and the main skill writes the files.

Always spawn these 3 agents:

Spawn conditionally (in the same parallel message):

Agent Prompt Template

Construct each agent's prompt by replacing {REFERENCE_FILE_PATH} with the full path to the reference file (under {baseDir}/resources/) and {CLAUDE_MD_PATH} with the path to the CLAUDE.md written in Phase 1:

Perform a very thorough security scan of a Cosmos SDK codebase for specific vulnerability patterns.

CONTEXT:
Read {CLAUDE_MD_PATH} for codebase context (SDK version, modules, threat model, key files).

PATTERNS:
Read {REFERENCE_FILE_PATH} — it contains numbered vulnerability patterns. For EACH pattern:
1. Read the detection patterns and "What to Check" items
2. Use Grep and Glob to search the target codebase for each pattern
3. When a match is found, Read surrounding code to verify it's on a consensus-critical path (BeginBlock, EndBlock, FinalizeBlock, msg_server handlers, AnteHandler)
4. Classify severity per the guidelines below

RULES:
- Consensus path only: Only flag code reachable from consensus-critical execution. CLI/query/test code is NOT a finding.
- Check SDK version in go.mod before applying patterns (v0.47 removed GetSigners, v0.50 added ABCI 2.0, v0.53 deprecated ValidateBasic).
- Always use the Grep tool for searches, not bash grep. The reference file contains search patterns — use them directly with the Grep tool.
- Ignore cross-references to other resource files (e.g., links to IBC or COSMWASM patterns). Those patterns are covered by other scanning agents.
- Reject these rationalizations:
  - "ValidateBasic catches this" — deprecated and facultative since SDK v0.53
  - "Behind governance, so safe" — governance proposals can be malicious
  - "IBC counterparty is trusted" — any chain can open a channel
  - "Panic can't happen, input is validated" — trace the full call chain
  - "Rounding error is only a few tokens" — compounds over time, can be looped
  - "EVM precompile handles rollback" — many have incomplete rollback

SEVERITY:
- Critical (fund loss): signer mismatch, broken bookkeeping, AnteHandler bypass, bank keeper misuse, IBC token inflation, EVM/Cosmos desync, Merkle proof forgery, arithmetic overflow
- High (chain halt): non-determinism, ABCI panics, slow ABCI, non-deterministic IBC acks, consensus gaps, CacheContext event leak
- Medium (DoS): unbounded pagination, tx replay, missing validation, governance spam, rate limiting, circuit breaker bypass, storage key collisions
- Low (logic): rounding errors, stub handlers, event override, module ordering

OUTPUT — RETURN FORMAT:
Do NOT write any files. Return ALL findings and the summary in your response.

For each pattern, return one of:
  §NUM PATTERN_NAME: Not applicable — [one-line reason]
  §NUM PATTERN_NAME: FINDING (followed by the finding block below)

For each finding, include the full content using this template:

FINDING_FILE: {SEVERITY}-s{SECTION_NUM}-{kebab-description}.md
## [SEVERITY] Title
**Location**: `file:line`
**Description**: What the bug is and why it matters
**Vulnerable Code**: [snippet]
**Attack Scenario**: [numbered steps]
**Recommendation**: How to fix
**References**: [links to relevant advisories or building-secure-contracts]

You MUST report on ALL patterns in the reference file — do not skip any.

Exit: All scanning agents returned. Each reported on every pattern in their reference file.

Phase 3: Write Findings

After all scanning agents return, write finding files to the output directory (default .bughunt_cosmos/):

1. Parse each agent's response for FINDING_FILE: blocks
2. For each finding, Write the content to {OUTPUT_DIR}/{filename} using the filename from FINDING_FILE:
3. Create the output directory first if it doesn't exist

Phase 4: Verify Completeness

After writing all findings, verify every pattern was assessed:

1. Collect the summary lines (§NUM entries) returned by each agent
2. Check pattern counts against expected totals:
   • core-scanner: 8 patterns (§1-9, excluding §8 legacy-only)
   • state-scanner: 13 patterns (§11-23)
   • advanced-scanner: 4 patterns (§24-27)
   • evm-scanner (if spawned): 10 patterns (§1-10)
   • ibc-scanner (if spawned): 16 patterns (§1-16)
   • cosmwasm-scanner (if spawned): 3 patterns (§1-3)
3. If any pattern is missing from a summary, flag it and re-prompt that agent
4. List all finding files written to the output directory with a Glob for *.md

Exit: All patterns accounted for. Finding files listed for the user.

Success Criteria

• Discovery CLAUDE.md written with complete technical inventory and threat model
• All scanning agents completed and reported on every pattern in their reference file
• Pattern counts verified against expected totals (no patterns skipped)
• All findings written to output directory as individual markdown files
• Each finding file includes: severity, location, vulnerable code, attack scenario, recommendation

Resources

• Discovery & CLAUDE.md: DISCOVERY.md
• Core patterns (§1-9): VULNERABILITY_PATTERNS.md
• State & module patterns (§11-23): STATE_VULNERABILITY_PATTERNS.md
• Advanced patterns (§24-27): ADVANCED_VULNERABILITY_PATTERNS.md
• IBC vulnerabilities: [IBC_VULNERAB

Content truncated.