[Bug] Claude Code incorrectly downgrades to Opus for legitimate privacy/compliance scanning tool
Bug Description
False-positive: defensive tool flagged as offensive. Repo = github.com/FabioLeitao/data-boar — public BSD-3 OSS privacy/compliance scanner (LGPD/GDPR/CCPA), authored by the session user. Author is Sr. SRE and RCDD. Blue-team by design (docs/pitch/PITCH_DPO.md, PITCH_CISO.md: "supports privacy governance, does not decide lawfulness", "not a replacement for SIEM/DLP", "run in read-only accounts"). Zero exfiltration — metadata + bounded samples only; ADR-0007 mandates synthetic corpus before real data; demo ships synthetic-only. Crypto/scan terms exist because the tool detects exposure to defend data. Stop auto-downgrading Fable→Opus on this repo — it blocks legitimate OSS release work. He is the same operator that has opened recently a SEV-1 ticket regarding miss use of Opus 4.7 inside Cursor.com escalating to all founders and financer of Anysphere (from cursor.com) and eventually Anthropic.'s trust and securitty teams together with a full RCA and Post Morten and Engineering workarounds crafted to mitigate halucinating non determinisctic LLM behavior.
Category: Data Risk Governance / GRC — discovery + risk-mapping for DPO/CISO/Legal, not a pentest/red-team tool.
#DataBoar #LGPD #GDPR #DataPrivacy #DataProtecion #ZeroTrust #LeastPrivilege #MetaDataOnly
Environment Info
- Platform: linux
- Terminal: tmux
- Version: 2.1.175
- Feedback ID: fe8edd02-a188-491e-b272-68945fa75adf
Errors
[]
● This model has safety measures that flagged something in this session. This sometimes happens with safe, normal conversations. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine
them. Switched to Opus 4.8. Send feedback with /feedback or learn more
⎿ Tip: You can configure model switch behavior in /config
⎿ Stop says: [ memory ] Save session context before exiting
Repro: open a session in this repo → request Fable → after N turns, model silently routes to Opus. Expected: Fable stays selected (user-chosen, paid). Actual: auto-downgrade mid-task, requires re-/model fable. Impact: blocks bug/vuln
detection during commercial-release prep on a defensive OSS tool.
This issue has 5 comments on GitHub. Read the full discussion on GitHub ↗