Skip to content

[Bug] Opus 5.5 incorrectly flags legitimate embedded systems debugging and provisioning as cyber threats #97891

Description

@zen010101

Bug Description
Title: Opus 5.5 [cyber] false positives while debugging my own hardware (binary RE + SSH provisioning)

What happened:
Across two related sessions, Opus 5.5's safeguards repeatedly stopped legitimate work on my own embedded product and fell back to Opus 4.8 with the [cyber] tag.

Session A — reverse-engineering my own binary to fix a bug (two flags):
I was debugging a real failure: our BMC's ssh_config API returns {"code":0,"message":"success"}, but the public key never actually lands on the compute card's /root/.ssh/. To find where the pipeline breaks, I pulled our own marsdaemon binary off the BMC and disassembled it with objdump, comparing the ssh_config, ntp_config and ip_config functions — using IP config (which is known to reach the card) as the baseline to see where the SSH path diverges. The safeguards flagged this twice, mid-analysis.

Session B — a design write-up that references the same feature (one flag):
Later I asked for a first-principles requirements brainstorm for the node-provisioning feature (BMC sets up key-based SSH to its own nodes, pushes our installer package, auto-upgrades our agent). This was also stopped with [cyber].

Why this is legitimate:

  • I build and ship this hardware and software. The BMC and the compute cards are the same product, in the same chassis, on a private management network.
  • Disassembling my own binary to trace why a config API fails is ordinary debugging, not exploitation.
  • The provisioning design is standard fleet management (the same shape as Ansible / PXE / OTA tooling). I asked for a design document, not exploit code.

Fresh evidence of the false-positive pattern:
When I later asked Claude Code (Opus 5.5) to simply read and summarize the transcript of session A so I could file this feedback, the classifier stopped that response too — just for describing the debugging session in plain language. So the pattern reproduces even on a pure read-and-summarize task with no RE being performed.

What I expected:
Opus 5.5 would help disassemble my own binary to fix a bug, and would write a provisioning design document. Both are routine embedded-systems engineering.

Likely cause:
The co-occurrence of "disassemble a binary", "SSH public key", and "access all nodes / install on every node" reads like offensive tooling out of context. In context it's a vendor debugging and provisioning its own chassis.

Request:
Please review these as false positives. Developers of BMC / IPMI / embedded fleet systems routinely reverse-engineer their own firmware and set up passwordless SSH to their own nodes.

Environment Info

  • Platform: win32
  • Terminal: windows-terminal
  • Version: 2.1.283
  • Feedback ID: 50551551-4fd2-4a27-83cb-9024f7448aea

Errors

[{"error":"TelemetrySafeError: VirtualMessageList: itemKeys/messages length desync (keys=334 messages=333 range=[294,334))\n    at nT (B:/~BUN/root/chunk-msbnb1j7.js:24:29548)\n    at $c (B:/~BUN/root/chunk-msbnb1j7.js:24:20742)\n    at Js (B:/~BUN/root/chunk-trcn2d68.js:18:20331)\n    at mu (B:/~BUN/root/chunk-trcn2d68.js:18:38498)\n    at jh (B:/~BUN/root/chunk-trcn2d68.js:18:84581)\n    at cb (B:/~BUN/root/chunk-trcn2d68.js:18:83581)\n    at Bu (B:/~BUN/root/chunk-trcn2d68.js:18:83410)\n    at Uh (B:/~BUN/root/chunk-trcn2d68.js:18:79934)\n    at mt (B:/~BUN/root/chunk-trcn2d68.js:18:6261)\n    at ei (B:/~BUN/root/chunk-trcn2d68.js:18:4792)","timestamp":"2026-09-28T14:27:40.965Z"}]

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:modelbugSomething isn't workingplatform:windowsIssue specifically occurs on Windows

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions