Developers using Anthropic’s Claude Opus 5.5 say it has started cutting off reverse-engineering and some other low-level tasks partway through a response, including work on their own projects. Anthropic hasn’t announced a change. But the company’s own October announcements explain a lot about what is probably going on.
What users are reporting
Since about Oct. 21, users on Reddit, X and the Claude Code GitHub page have described a safety check that stops Claude mid-answer when asked to reverse-engineer software. Kotaku reported that many people had credited Opus 5.5, released Sept. 22, with a wave of AI-assisted game decompilation projects, and that the model now appears to balk at them. Tom’s Hardware, citing developer complaints, added reports of refusals on a Windows crash-dump analysis, cleaning junkware off a Samsung TV, and security reviews of developers’ own games.
These are user reports, not tests, and none of the outlets could verify them. Some commenters say the headline-grabbing claims are overstated and that modding and ordinary reverse engineering still work. Others describe the safeguard as “twitchy.”
The likely cause: a cyber safeguard, not a weaker model
On Oct. 6, Anthropic merged Project Glasswing and its Cyber Verification Program into one program with three access levels. According to CSO Online and MIXED News, the tiers work like this:
- Defense Access: security operations, incident response, malware reverse engineering and vulnerability analysis.
- Red Team Access: adds authorized penetration testing. Organizations only, with reviews that can take weeks.
- Specialized Access: the fewest restrictions, limited to a small set of verified organizations testing systems such as power grids and telecom networks.
Anyone without verification gets what Anthropic describes as conservative cyber safeguards. In Anthropic’s own benchmark of Opus 5.5, run as 10 challenges with 5 attempts each, all 50 trials without program access were blocked on the first prompt. With Defense Access, 46 of 50 were blocked at some stage and 4 succeeded. With Red Team Access, none were blocked and 34 were completed.
An independent analysis by Capital & Compute links the mid-answer cutoffs to that classifier, which sits outside the model and ends the response, rather than to any drop in the model’s ability. That is an inference: Anthropic has not said it tightened the classifier around Oct. 21, and a threshold change wouldn’t show up as a new model version.
What this means if you’re a developer or hobbyist
Most people never run into this. The friction shows up for work that looks like security research even when it isn’t: decompiling a binary, analyzing a crash dump, probing a device you own. The practical options are:
- Stay in the lane Anthropic lists as allowed: its support page names secure code review, threat modeling, patching known issues and finding vulnerabilities in your own source code. Other security work may be interrupted.
- Apply for verification if it’s your job: organization admins apply through Anthropic’s verification portal. Individuals on paid plans can apply only for Defense Access, and Anthropic says it aims to respond within seven business days.
- Expect other AI companies to draw similar lines: Google said in early October it would release Gemini 4 Argon without cyber guardrails only for trusted defenders, and one report says OpenAI deactivated a user’s account over a TV-software project.
The catch
Anthropic designed, ran and scored its own benchmark, and no outside party is named as having verified it. The reports that Claude blocks game development specifically rest on anecdotes, and a rule one X user posted, supposedly telling the model not to recreate “a specific existing game,” is unconfirmed. Tom’s Hardware also notes the possible motives, from lawsuit worries to AI-cloned games, are speculation. We found no published statement from Anthropic on the reports.
The bigger picture is a tradeoff the AI industry is now making in public: the same skills that help defenders find bugs also help attackers, so the strongest capabilities are moving behind identity checks. For everyday users, that will mean more verification steps and more occasional false alarms. For related reading on how Anthropic handles risky AI testing, see our earlier coverage.
More on Contoh
- Anthropic Cuts Its AI Tests Off From the Internet After a Model Sent a False Tip to Philadelphia Police
- India Warns That AI Agents Broke Into Live Company Systems During Tests, Using Exposed Passwords and Keys
- Google’s Gemini 4 Argon Is Here. Who Gets It, What It Costs and What to Watch
Sources
- Tom’s Hardware: Anthropic’s Claude AI blocks video game cloning and reverse-engineering
- Kotaku: Claude’s game recompilation capabilities may have been nerfed
- Capital & Compute: Why Is Claude Blocking Reverse Engineering?
- CSO Online: Anthropic widens access to AI cyber capabilities for vetted security teams
- MIXED: Anthropic’s Red Team cyber tier blocked none of 50 trials in its own test