Anthropic Releases Claude Opus 4.7 with Cybersecurity Safeguards

Gate News message, April 17 — Anthropic announced Claude Opus 4.7, an AI model featuring safeguards designed to block high-risk cybersecurity requests. The release builds on Claude Opus 4.6, introduced in February, and precedes a broader rollout of Anthropic’s more advanced Mythos-class models. Mythos Preview, introduced earlier this month, remains more capable than Opus 4.7 but is currently available only to a limited set of companies through Project Glasswing, Anthropic’s cybersecurity program.

Opus 4.7 delivers significant improvements in agentic workflows and vision capabilities. Autonomous coding performance rose to 64.3% on SWE-bench Pro, a widely used benchmark for software engineering tasks, from 53.4% in the prior version. Image handling now supports over three times the resolution, with visual-acuity accuracy improving from 54.5% to 98.5% on standard tests. Intentional cybersecurity limitations were implemented during training; on CyberGym, a cybersecurity performance benchmark, Opus 4.7 scored 73.1%, below Mythos Preview’s 83.1%.

Anthropic restricted Mythos Preview following pre-release testing that uncovered thousands of high-severity vulnerabilities, including a flaw in OpenBSD undetected for 27 years. Security researchers conducting vulnerability research, penetration testing, or red-team exercises can apply to Anthropic’s Cyber Verification Program for defensive access without default refusals. Feedback from Opus 4.7’s safeguards will inform future model releases.