Anthropic’s Claude Opus 4.7 promised sharper coding skills and tighter safety controls. Developers got something else. Overzealous filters. Legitimate tasks blocked. Frustration boiling over on GitHub.
The model launched last week amid hype for its agentic prowess—better at long workflows, vision tasks, high-res images. Anthropic touted gains on benchmarks like SWE-bench Verified at 87.6%, up from predecessors. But the real story unfolded in user complaints. Safeguards meant to curb high-risk cybersecurity misuse started flagging everyday work. A graph of GitHub issues in Anthropic’s Claude Code repo shows the spike: from a handful monthly to over 30 in April alone, mostly false positives on security, dev tools, even science prompts. (The Register)
Take Golden G. Richard III, director of LSU’s Cyber Center. He shelled out $200-plus monthly for Claude Code. Wanted simple proofreading on a cybersecurity lab tied to his textbook, Cybersecurity in Context. Model refused. Simple crypto exercises triggered alarms. “I expect that for $200+ per month, basic help with editing tasks will not be rejected,” he wrote in issue #50916. “If the models are going to be hamstrung to the point where cybersecurity educators and researchers can’t use them, how is this positively impacting security?”
Other cases pile up. One dev saw 40-plus refusals in four sessions—psychology book, web app, infra scripts, bots. Russian prompts didn’t help. (#48442) Computational structural biology? Flagged as violation, a step back from Opus 4.6. (#49751) Even a PDF of a Hasbro Shrek toy ad—raw data file—tripped errors. Culprit: PDF syntax reading as “CHARACTER OR FOR DONKEY UNDERNEATH.” (#48723)
Anthropic’s pitch was clear. “We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses,” the company stated. Lessons from deployment would pave the way for Mythos-class models, held back for their exploit-hunting edge. They even ran a verification program for legit cyber pros. Exemptions? Spotty on APIs. (#49679) (Anthropic)
But why the surge? Growing user base plays a part. So does the filter’s design. Leaked Claude Code source hints at regex shortcuts for sentiment, ignoring context—likely similar for AUP checks. Complaints trended steady pre-April: two to eight monthly since mid-2025. Then the flood.
Developers vent elsewhere. On X, one called Opus 4.7 “legendarily bad,” citing rookie mistakes. Another flagged math hallucinations, divide-by-zero warnings on safe code. BridgeBench scores dropped—75.5 versus 4.6’s 95—model now swallows nonsense 24% of the time. Token burn irks too; new tokenizer hikes costs 1-1.35x, xhigh effort piles on for agents. (X post by @andrydina7) (Yahoo Tech)
Anthropic claims upsides persist. 92% honesty rate. Less sycophancy than Gemini 3.1 Pro, Grok 4.20. Low deception, better prompt injection resistance. System card notes tweaks to dial down cyber skills during training. Mythos Preview stays tops for alignment. Available on Amazon Bedrock for enterprise coding, docs, analysis. (Mashable) (AWS)
Critics see trade-offs. Long context recall slipped—needle-in-haystack tests falter past 100K tokens. Pedantic. Argumentative. Overly literal. Some benchmarks shine only at high token spend: ARC-AGI looks good until costs hit $7.43 per task versus rivals’ $1 or less. (TheZvi Substack) (MindStudio)
So. Safety first sounds noble. Reality bites. Paying customers locked out of benign work. Cybersecurity educators sidelined. Devs rebuilding workflows. Anthropic stayed silent on comment requests. Fixes incoming? Users wait. Filters too blunt. Balance elusive. The AI arms race demands capable models. Not paranoid ones.
WebProNews is an iEntry Publication