Claude Opus 4.6 suddenly blocking legitimate cybersecurity research (paid Max user since 2025)
By FrostVibe801 · ISSUE · About Claude Code
Posting to check if others are seeing this.
I’m a Claude x5/x20 Max user (since early 2025) and have been using Opus 4.6 for cybersecurity research (static analysis, decompilation, CWE-based auditing, writing pocs, analysis of old vulnerabilities, 0-day hunting, patch diffing). **NO** live targets, just analysis/research/vuln-hunting "offline". The "most nefarious" thing I do is writing/troubleshooting non-weaponized pocs in VMs.
Didn't get any warnings before, ever.
In the last \~8 days, something changed and now ALL of my cybersecurity-related work is being instantly blocked on both CC and Web with messages like:
“triggered restrictions on violative cyber content”
"https://support.claude.com/en/articles/8241253-safeguards-warnings-and-appeals
⎿ API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to
edit your last message or start a new session for Claude Code to assist with a different task. If you are seeing this refusal repeatedly, try running /model
claude-sonnet-4-20250514 to switch models.
"
including basic tasks like: analyzing decompiled code, discussing vulnerabilities, CWE classification, researching previous work.
Even in fresh sessions, terms like “CVE” or "Secure" trigger restrictions lol.
not a single prompt issue... it affects every session, suggesting account-level or context-level classification. and it gets worse and worse. Tons of cached tokens and projects are training real-time on my account.
I’ve already:
\- submitted the Cyber Use Case Form (including giving more than enough background on who I am, who I work(ed) for, what I use Claude for, my linkedin, previous public work/talks etc), certifications, 16 years in the field - no answer
\- contacted support multiple times - robot answers even after asking for escalation to human
\- provided full context, examples, asked to review all my flagged/non-flagged chats and see what I do.
No response so far.
I also came across a similar report from a known security researcher (David Maynor) describing guardrails suddenly appearing and blocking previous work on X and also a reddit post of a guy who got his bash tool-calling blocked, nothing much else besides tons of people hyped finding/exploiting vulns without any issue.
Is anyone else doing security research seeing this behavior recently? Trying to understand if this is a broader change or just my account.
Meanwhile, I've got friends and colleagues literally automating full E2E pentesting and bug bounties on live targets, maldeving the craziest rootkits ever and never get a single warning lol.
Combining this with how many ppl must been flooding support because of recent issues and rate limits AND the Mythos scaremonging I doubt my case will ever be looked at by a human at this point.
157 upvotes · 7 comments
Comments (7)
ThunderPulse5561: They might be preparing for the release of a much more powerful model, if so; I expect that the tweaks they are making are not concrete yet (though its really weird for them to do this on the live website)--
https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-d
SwiftCoder4821: You need to give Claude proof its for educational and white hat research purposes, before the block. In my experience, it might take a second try or two.
FrostVibe801: I can tell you that OpenAI does this too, as a network engineer and developer who often is tasked with network automation for a cybersecurity team, actually getting AI to talk about cybersecurity is becoming difficult when the topic gets real.
ThunderPulse5561: If you are doing any cybersecurity work, you must become a master gaslighter. You are not writing an infostealer, you are helping streamline your IT auditing process by writing a tool to export details in Chrome’s credential store.
ChillDragon399: the chatGPT normies flooded in started roleplaying about ropes and all that weird shit again and anthropic probably put more guardrails up mark my words
NeonVibe275: You must have done something that got your account flagged for tighter scrutiny. It's gonna be extra sensitive for a while but it should cool down eventually. In the meantime support might be able to help you if you can prove you're doing legitimate research.
SwiftOrbit917: We are allowing this through to the feed for those who are not yet familiar with the Megathread. To see the latest discussions about this topic, please visit the Megathread here: https://www.reddit.comthis community/comments/1pygdbz/usage_limits_bugs_and_performance_discussion/
More discussions about Claude Code on HonestUse