IDA and readable code
Reopen an IDA database, inspect annotated pseudocode and follow the recovered scope.
AI-ready binary analysis
Make your AI reverse-engineering agent useful when protected code stalls.
Devirt is the recovery and evidence layer for protected binaries. Use REA, IDA, Ghidra, Binary Ninja or your own model for the investigation; Devirt returns checked artifacts and a ledger your agent can read, cite and continue from.
Customer-controlled models. Read-only MCP access. Files you own or are authorized to analyze.
Definition
What is AI reverse engineering?
AI reverse engineering is the use of large language models and AI agents to analyze compiled software without its source code. An agent drives tools such as IDA Pro, Ghidra or Binary Ninja, ranks functions, names symbols and explains what the code appears to do. It helps most where the decompiler output is readable.
On virtualized code, AI reverse engineering stalls. The protected function becomes bytecode run by an interpreter, so the agent sees a dispatcher and handlers instead of the logic. A 2026 review of agentic reverse-engineering systems lists obfuscation among the cases where they still fail, and our reverse engineering MCP guide shows where that happens in an agent workflow.
Devirt does not replace the agent. Devirt recovers the protected scope and places every in-scope function in one of four proof states, so the agent continues from checked facts instead of a confident guess.
The agent loop
Keep the investigation in your tools.
- Inspect locally.Ask your agent to map functions, calls, strings and the blocked question.
- Assess the boundary.Upload the authorized file and receive a feasibility decision, scope and quote.
- Recover the accepted scope.Devirt handles the protected boundary and records what each check established.
- Continue with evidence.Your MCP client reads the result and brings the artifacts back into the investigation.
AI-readable evidence
A model gets boundaries, not a confident guess.
Every accepted Devirt result includes machine-readable metadata for the file, exact build, scope, assumptions, proof state, artifacts and remaining questions. The coverage ledger keeps facts, inferences and unknowns separate so an agent can cite the evidence instead of turning an unresolved path into a claim. Our research explains how the four proof states are decided.
ProvenThe accepted check supports the stated behavior within the declared scope.
Runtime-dependentThe result needs runtime evidence before an agent should treat it as settled.
UnresolvedThe available evidence did not establish the behavior.
Permanently boundedThe question is outside the accepted model or artifact boundary.
Agent-readable metadata does not promise original source code. Names and structure removed by compilation or protection may remain unavailable.
Tool-ready handoff
Use the result where your team already works.
Structured workflows
Consume API JSON and proof metadata from REA, another MCP client or your own model pipeline.
Human review
Read the report, assumptions and unresolved questions before relying on a conclusion.
Agent Starter path · for individual researchers
Start small. Upgrade only when the file needs it.
Community Triage
Bounded file and protection indicators before you commit.
Create an accountTriage Plus
Downloadable JSON evidence for your local agent workflow.
See Triage PlusGeneral Analysis
Ordinary analysis or an eligible small VMProtect/Themida scope with IDA and proof JSON.
Assess for GeneralPower
One General allowance, five Triage Plus scans, read-only MCP access and 10% off eligible recovery.
Choose PowerBroader protected recovery starts with a free assessment and a written quote from $399.
You remain in control.
Devirt’s MCP connector is read-only. Upload, authorization, scope approval and payment happen in your account, not through an agent tool. Bring your own model and revoke the key from workspace settings whenever you choose.
Questions agents and researchers ask
Clear limits before you upload.
Is Devirt an AI model?
No. Devirt performs scoped recovery and verification; your chosen model or agent interprets the evidence.
Can REA read the result?
Yes, through your agent. You register Devirt’s read-only MCP connector in the same MCP client as REA, and it exposes account-scoped assessments, jobs, results and structured ledger metadata beside your other MCP tools.
Can AI reverse engineer protected binaries?
An AI agent can map a protected binary, but a virtualized function stays bytecode until it is devirtualized. The agent can identify the protector, trace unprotected code and describe the VM interpreter. Devirt performs the devirtualization for files you own or are authorized to analyze (see VMProtect devirtualization and Themida devirtualization) and returns each recovered function with its proof state.
Is AI reverse engineering legal?
Using AI does not change the legal question: whether reverse engineering is allowed depends on the jurisdiction, the software license and the purpose. Devirt accepts only files you own or are authorized to analyze, and it does not remove DRM, bypass license checks or crack software. Read the acceptable use policy before uploading.
What if the binary is too complex?
We can request more evidence, quote a different scope or decline it. You approve the written offer before paid work begins.