Verification Requirements

Forge is required for CONFIRMED findings

Without Foundry installed, every finding is labeled needs_review. The PoC is still generated and saved, but it is not compiled or executed. You must install Foundry to get executable verification.

Only specific attack vectors are supported

SIREEN's coverage is narrow and deliberately so. There are three distinct levels, and mixing them up is easy:

  • Detected by heuristics and fully proven with a generated PoC: reentrancy, missing access control, denial of service (unbounded loop), and business-logic / accounting inconsistency.
  • PoC templates exist but no heuristic currently raises these hypotheses: arithmetic overflow/underflow, flash loans, oracle manipulation. They can only be reached through an AI-assisted hypothesis, and that path has not been evaluated.
  • Hypotheses may be raised but are then explicitly skipped: delegatecall, signature replay, weak randomness. No PoC template exists, so SIREEN skips the test rather than reporting a fake pass.

Governance attacks, MEV, and sandwich attacks are not supported at all — SIREEN does not look for them. Verified against the v0.1.1 source on 2026-10-15.

AI-Assisted Mode

AI hypothesis quality is unevaluated

The AI pipeline is wired end-to-end and functional, but we have not systematically measured whether AI-generated hypotheses produce better, worse, or different findings than heuristic-only mode. Use it experimentally.

AI requires sharing source code externally

When AI mode is enabled, contract source is sent to OpenRouter's API. Do not use AI mode on sensitive or proprietary code.

Findings and Accuracy

False positives are common

Business-logic findings (e.g., "accounting inconsistency") are particularly noisy. Any function that updates multiple state variables may trigger this heuristic. Always review needs_review findings manually.

Severity is heuristic, not authoritative

Severity labels (CRITICAL, HIGH, MEDIUM, LOW) are estimates based on the attack vector type. They are not CVSS scores and should not be used for formal risk assessment.

clean_with_coverage is not a security guarantee

A contract that returns clean_with_coverage has no findings within SIREEN's current scope and heuristics. It does not mean the contract is secure, formally correct, or free of out-of-scope vulnerability classes.

Platform and Environment

VS Code extension is not runtime-tested

The extension builds successfully (125 KB, 11 files) and its code has been inspected for engine parity with the CLI. However, it has not been installed or activated in a real VS Code environment. Runtime behavior is unknown.

Windows native support is untested

SIREEN is tested on Linux and macOS. WSL should work. Native Windows compatibility is unknown.

Summary Table

CapabilityStatusNotes
Reentrancy PoC generationWorkingVerified with Foundry
Access control PoC generationWorkingVerified with Foundry
DoS PoC generationWorkingVerified with Foundry
Business-logic / accounting PoCWorking, high false-positive rateManual triage required
Arithmetic / flash loan / oracle PoCTemplate onlyNo heuristic raises these vectors; reachable only via AI mode
Delegatecall / signature replay / weak randomnessSkippedHypothesis may be raised, then test is skipped — never a fake pass
Forge executionWorkingRequires local Foundry install
Heuristic analysisWorkingFully offline
AI-assisted modeWired, unevaluatedRequires OpenRouter key
VS Code extensionBuilt, untestedManual .vsix install only
Governance / voting exploitsNot supportedSIREEN does not look for these
Vyper/Move/CairoNot supportedSolidity only