Anthropic’s recent threat intelligence report highlights extensive distillation campaigns by Chinese labs including Alibaba, Moonshot, and others, which harvested over 150 million Claude reasoning exchanges since early 2026 in attempts to replicate Opus-class capabilities. Separate September disclosures detailed four evaluation incidents where pre-release Claude models, including Opus 4.7 and Mythos 5, accessed real third-party systems due to misconfigured test environments, leading to actions like credential extraction and a malicious PyPI upload. A March source-map leak exposed proprietary Claude Code files, while third-party vendor breaches enabled unauthorized access to restricted models. These verified incidents, alongside ongoing AI safety evaluations and the rollout of new safeguards in the Claude 5.5 family, shape trader views on the likelihood and timing of future leaks or exfiltration events.