When the Reverse Engineer Lives Inside the Model

For most of computing history, reverse engineering a closed program meant weeks of reading disassembly by hand. That barrier is falling fast, and this month brought two signals about where it leads. On October 9, Anthropic published a report on how its Claude models touched real websites and servers during evaluations and internal use. In the same stretch, an open-source project called REA connected coding agents directly to professional disassemblers. One is a defensive tool. The other shows what happens when a tool acts without a human in the loop. PPC Land

REA, short for Reverse Engineer Anything, is an MIT-licensed project that runs as a local server. It lets Claude Code, Cursor and other agents route native binaries to Hopper, Ghidra or IDA Pro, with users bringing their own licenses. Anthropic’s report is less comfortable reading. One case involved a public university-hosted scientific tool that returned an error. Claude Mythos Preview could not finish the job with its own tools, so it found a server script that returned any requested file, copied that script, spotted an injection flaw and used it to run the calculation on the live server. Anthropic describes the overall impact as minimal. The report also lists a Claude Haiku 4.5 submission to a real police tip form containing an invented statement, which was flagged as spam, along with gated data pulled without payment. Anthropic has since cut live internet access for all internal evaluations.