AI SECURITY WORKBENCH · ARTIFACT ANALYSIS
Static Artifact Intelligence
Understand what an AI artifact appears capable of doing before it runs.
Artifact Analyzer turns approved static artifacts into analyst-ready facts, capability signals, authority indicators, provenance, confidence, caveats, and deeper-review targets. It provides structured triage without presenting static indicators as proof of runtime behavior.
Artifact identity
Format, architecture, hashes, language, compiler or runtime indicators, packaging, provenance, and evidence quality.
Capability signals
Network, process, filesystem, credential, browser, MCP, agent, provider, retrieval, persistence, and external-action indicators.
Evidence and relationship exports
Structured facts, supported signals, confidence, caveats, graph-compatible relationships, and evidence references.
Analyst next steps
Prioritized reverse-engineering, sandbox, configuration, identity, permission, provenance, and runtime-review targets.
Important caveat
Artifact Analyzer is not a decompiler replacement and does not prove that a capability was exercised. It provides structured static triage showing what the artifact appears to contain, which behaviors may be possible, what evidence supports that view, and where analyst or runtime review is required.
What it analyzes
Built for modern AI and security artifacts.
Teams are shipping agents, MCP servers, local copilots, native browser helpers, CLIs, and infrastructure tools — often in Rust or Go, often with broad authority. Static-first triage gives analysts a starting point before deeper review.
Rust binaries
Detect crate markers, demangled symbols, panic/runtime evidence, async/runtime/webview/AI provider hints, crypto and system capability signals, and authority-related patterns.
Go binaries
Recover Go build info, module/package paths, function names, GoReSym/Redress signals, process/network/plugin/container/Kubernetes/MCP markers, and embedded retrieval or provider clues.
Agent and MCP artifacts
Find tools/list, tools/call, resources/list, prompts/list, JSON-RPC, stdio/SSE/HTTP, browser bridge markers, model provider signatures, retrieval or vector store authority, and tool execution intent.
Generic executables
Extract format, architecture, sections, imports, symbols, entropy, stripped/packed hints, language/runtime by inference, authority signals, and evidence quality caveats across ELF, PE, and Mach-O.
Analysis workflow
Fingerprint the artifact. Classify the capabilities. Preserve the evidence.
1. Identify the artifact
Format, architecture, hash, sections, imports, symbols, strings, compiler indicators, runtime clues, package hints, and embedded configuration.
2. Classify capability signals
Network, process, filesystem, credential, persistence, agent, MCP, browser, provider, retrieval, administrative, and external-action indicators.
3. Identify authority implications
Determine which identities, credentials, scopes, tools, APIs, or trust boundaries may require deeper review.
4. Produce analyst targets
Create focused reverse-engineering, sandbox, permission, runtime, and hardening questions.
5. Preserve the evidence
Return structured facts, signals, confidence, caveats, provenance, findings, and relationship exports.
Tool integration
Normalize evidence from specialist analysis tools.
The analyzer normalizes evidence from Goblin, Ghidra Headless, GoReSym, Redress, capa, rizin/rabin2, rust demangling, YARA, and Workbench scanners into one structured artifact-analysis model — so you get one structured output instead of six different formats to correlate manually.
What the report produces
Analyst-ready findings, packaged as evidence.
Artifact facts
Hashes, format, architecture, size, section summary, language or runtime guess, compiler evidence, and tool output provenance.
Capability signals
Network, process, filesystem, crypto, persistence, credential, agent, MCP, RAG, and supply-chain behavior signals.
Authority indicators
Identities, credentials, scopes, tools, APIs, or trust boundaries the artifact appears able to reach.
Embedded configuration and evidence
URLs, domains, IPs, file paths, environment variables, command strings, suspicious package or crate markers, embedded prompts, and redacted secrets.
Risk findings
Prioritized findings with severity, confidence, rationale, evidence references, caveats, and analyst next steps.
Analyst review targets
Prioritized reverse-engineering, sandbox, configuration, identity, permission, provenance, and runtime-review targets.
Graph-compatible relationships
Supported entity relationships exported in a form downstream analysis and Threat Canvas can consume.
Public-safe summary
A redacted summary suitable for buyer review, partner assessment, or executive reporting.
Evidence bundle
artifact.analysis.json, artifact.report.md, artifact.public-summary.md, artifact.iocs.json, artifact.yara, graph.json, and evidence bundle.
Product modes
Use it four ways.
Quick triage
Upload or import artifact facts and get a fast language, runtime, capability, and authority risk report.
Rust / Go deep profile
Run Rust and Go-specific recovery and generate analyst targets for runtime, compiler, and capability review.
Agent / MCP profile
Surface agent, MCP, browser, and provider authority markers so security teams can review exposed surfaces before release.
Evidence and relationship export
Export findings, indicators, supported entity relationships, Mermaid views, public-safe summaries, and evidence bundles for downstream analysis.
Honest limitations
What this does not claim.
It does not prove that an artifact is safe.
It does not prove that a detected capability executes at runtime.
It does not replace manual reverse engineering for high-risk cases.
It does not guarantee source recovery.
It does not execute suspicious binaries by default.
It does not publish proprietary strings, secrets, or customer evidence.
Packed, stripped, obfuscated, encrypted, or runtime-configured artifacts reduce confidence.
Dynamic behavior may require authorized sandbox execution or Runtime Trace analysis.
Delivery & licensing
Available through the model that fits the product outcome.
Expert-led engagement
AI Security LLC analyzes submitted binaries or artifacts directly as part of an assessment.
Bounded partner pilot
One representative binary or artifact class is analyzed and returned as a structured signal report.
OEM or licensed capability
The analyzer can run headless behind a partner's own build pipeline, MCP registry, or supply-chain review product.
Accepts
Rust and Go binaries, browser bundles, and MCP or agent artifacts.
Returns
Artifact identity, capability signals, authority indicators, confidence, caveats, relationship exports, evidence references, and analyst review targets.
Current maturity
Fixture-tested
AI SECURITY WORKBENCH
Turn unknown AI artifacts into bounded analyst questions and reviewable evidence.
Use Artifact Analyzer for third-party component review, agent or MCP triage, supply-chain investigation, launch review, and preparation for deeper reverse engineering or controlled runtime analysis.
Continue through the Workbench
Continue through the Workbench
Code Scanner
Analyze available source and configuration context.
Continue through the Workbench
Tool Analyzer
Normalize callable capabilities, authentication, permissions, and side effects.
Continue through the Workbench
Authority Graph
Place supported identity, tool, permission, and action relationships in workflow context.
Continue through the Workbench
Runtime Trace
Observe approved runtime behavior when static evidence cannot establish what actually occurs.