Summary
Given an unpacked router firmware image, an agent mapped the filesystem, summarised the services and ranked binaries by risk faster than I would have by hand. It also invented a hard-coded key in a function it had not read. Splitting the work into steps it could check made the difference.
Background
Firmware triage is a long tail of small questions: which services listen, which binaries parse untrusted input, which scripts shell out. Each is easy to answer and tedious to answer a hundred times, which is where an agent should help.
Method
I gave the agent a read-only view of the extracted filesystem, a disassembler it could script, and a rule that any claim about behaviour had to quote the code or the command output that supported it.
Findings
Enumeration and summarisation were reliable. Reasoning about stripped binaries without running them was not: when it had to guess, it guessed fluently. Claims that came with a quoted decompiler snippet were right far more often than claims that did not.
Takeaways
Use agents for breadth and make them cite evidence for depth. If a statement cannot be checked in one command, treat it as a hypothesis.