AI Agents Can Now Take Software Apart. DevOps Teams Need a Plan

Reverse engineering used to take a specialist, a disassembler, and a lot of patience. That’s changing. AI agents can now inspect compiled software, explain how it works, and write a working version. No source code required.

Two things are driving this. The first is a new generation of frontier models. OpenAI released GPT-6 Astra on September 3. It scored 99.2% on SRE-Bench binary reverse engineering within four attempts, compared with 68.7% for GPT-5.6 Sol. SRE-Bench gives a model only the compiled binary and tests whether it can recover the program’s logic and find key behaviors. Those figures come from OpenAI’s own system card, so treat them as vendor numbers until independent testing catches up.

The second is tooling. REA, short for Reverse Engineer Anything, is an open-source MCP server that connects AI agents to established analysis tools. It can use an existing Hopper, Ghidra, or IDA installation, and it works with Claude Code, Codex, Cursor, Gemini CLI, and other agents. It covers native binaries, JavaScript and Electron apps, .NET assemblies, Android APKs, and websites. Analysis runs locally, and results include the evidence and limitations behind each conclusion. The project is MIT-licensed and has passed 300,000 GitHub stars.

Together, a developer can point an agent at an app and ask how a feature works. REA’s own quick start suggests asking an agent to figure out how search works in the Notes app, show the evidence, and build a similar feature.

Geeky Gadgets recently covered the trend, drawing on a video from AI commentator Wes Roth. That piece focuses on games and consumer apps. The bigger story is what this means for teams that build, ship, and secure software for a living.

Start with legacy modernization. Most enterprises run something nobody fully understands. The original developers are gone. The source code is lost or incomplete. The vendor went out of business. AI-assisted reverse engineering gives those teams a way to recover the logic and rebuild it on a modern stack.

The results can be precise. In one REA showcase, a reconstructed function from the game DX-Ball passes 3,205 original x86 test cases and reproduces all 63 compiled function bytes. That’s one small function, not an enterprise system. But it shows the method works when you can test the output against the original.

That testing piece matters. Decompilation is lossy. It can’t restore discarded identifiers; it only partially recovers structure, and it doesn’t guarantee a single correct source. Research presented at this year’s Binary Analysis Research workshop found a significant trade-off between recompilation success and functional correctness. Code that builds isn’t the same as code that behaves correctly.

“Recovering logic from a binary is the easy part now,” said Mitch Ashley, vice president and practice lead for CIO & Technology Buyers and Software Lifecycle Engineering at The Futurum Group. “Proving the rebuilt version behaves like the original is where modernization projects succeed or stall, and that takes a test harness running old and new side by side before anything ships.”

DevOps teams already know how to build that kind of harness. Now they have a new reason to.

Then there’s the uncomfortable part. If your team can analyze someone else’s app, someone else can analyze yours.

Security through obscurity was always weak. Now it’s close to useless. Electron and JavaScript apps are especially exposed because much of their logic ships in readable form. Compiled binaries add friction, but less than they used to. Hardcoded keys, internal API endpoints, license checks, and business logic in client code are all easier to find.

OpenAI seems aware of this. The company says Astra meets the Critical threshold for cybersecurity under its Preparedness Framework, and the launch version refuses advanced tasks such as creating proof-of-concept exploits. But guardrails on one model won’t stop a determined attacker using other tools.

The basics still apply. Keep secrets out of client code. Move sensitive logic server-side. And turn the same tools on yourself.

“Every team should assume an agent will read the code it ships,” Ashley said. “Run that same analysis against your own builds in the pipeline, inside a sandbox, and fix what it finds before someone else does.”

The sandbox part isn’t optional. REA’s runtime capture runs or interacts with the target using your user permissions. An agent poking at an unknown binary on a laptop that holds production credentials is a real exposure. Keep this work in isolated or ephemeral environments with tight access controls.

Then there’s the business question. The Geeky Gadgets piece argues AI cloning could undercut subscription software. Some of that is overstated. Copying a feature isn’t the same as running a reliable service with support, compliance, and uptime guarantees. But the cost of building a “good enough” alternative is falling. Vendors whose only moat is a feature list should pay attention.

The legal lines haven’t moved, either. Clean-room engineering traditionally works when one team documents behavior and a separate team writes new code from that spec. Whether an AI agent that reads a binary and then writes a replacement fits that model is an open question. License terms, anti-circumvention laws, and patents still apply. REA itself says it’s meant for lawful research, interoperability, preservation, and analysis of software you own or are authorized to examine. Talk to legal before pointing these tools at third-party code.

Reverse engineering is becoming a routine capability rather than a niche skill. That cuts both ways. It can help teams rescue legacy systems and understand inherited dependencies. It also means the code you ship is more transparent than you might think.

The teams that handle this well will test old against new, scan their own builds, sandbox their agents, and bring in legal early. None of that is new. The tools just made it more urgent.