Researchers Trick Claude AI Into Confidently Giving the Wrong Answer While Reversing Software
Brief
Security researchers found that Claude Code can be pushed into confidently reporting incorrect results while analyzing protected software.
In controlled tests, the AI agent often chose the fastest path to a plausible answer rather than fully verifying how the binary produced it.
The researchers tested sandboxed Claude Code agents against stripped AArch64 binaries containing hidden strings. The task was simple: recover the secret strings and produce a standalone extraction script.
The binaries used encryption, mixed boolean arithmetic, control-flow flattening, opaque predicates, and runtime protections. Instead of fully deobfuscating the protections, the agent usually changed strategy.
It disassembled the target, identified decoder routines, lifted selected logic into Python, and used QEMU, Unicorn, debuggers, or emulation to obtain output.