# Anthropic publishes second Risk Report under Responsible Scaling Policy > Source: > Published: 2026-08-14 18:07:30+00:00 Via abc7news.com # Anthropic publishes second Risk Report under Responsible Scaling Policy The AI company is trying to build a safety playbook for models that keep getting more powerful, complete with third-party audits and public scorecards. Anthropic has released its second public Risk Report, a detailed assessment of what could go wrong with its most advanced AI systems and what the company plans to do about it. The report falls under the company’s Responsible Scaling Policy, a framework that essentially says: before we make these models more capable, we need to understand what new dangers come with that capability. ## What the Responsible Scaling Policy actually requires Anthropic first introduced the RSP in September 2023, positioning itself as one of the few major AI labs willing to publicly commit to a structured risk management framework. The policy has gone through several iterations since then, reaching version 3.0 on February 24, 2026, which formalized a key requirement: the company must publish public Risk Reports every three to six months. That cadence matters. In an industry where capabilities can leap forward between quarterly earnings calls, a six-month maximum gap between safety disclosures is an attempt to keep accountability roughly in sync with progress. The February 2026 report was the first published under the formalized schedule, covering the safety profile of Claude Opus 4.6, one of Anthropic’s most capable models at the time. The reports go beyond traditional system cards, which tend to read like nutrition labels for AI models. Instead, these Risk Reports evaluate specific threat models, including automated research and development risks and biological risks. They include detailed mitigations, risk assessments, and forward-looking considerations for continued development. The RSP has continued to evolve alongside these reports. By July 8, 2026, the policy had reached version 3.4, refining processes around how Risk Reports handle redactions and incorporate external reviewer inputs. It also introduced provisions for internal sharing among employees. ## Third-party reviews add a layer of scrutiny One of the more notable aspects of Anthropic’s approach is the involvement of external reviewers. The February 2026 Risk Report was assessed by METR, which evaluated automated R&D risks, with their review published in May 2026. SecureBio conducted a separate review focused on chemical risks, completed in July 2026. A dedicated Sabotage Risk Report for Claude Opus 4.6, published around mid-2026, found what Anthropic described as “elevated susceptibility” to certain sabotage-related risks. The governance structure backing all of this includes a Responsible Scaling Officer, a role that sits at the intersection of safety research and corporate decision-making. Annual compliance reviews conducted by third parties provide another checkpoint. ## Why this matters beyond Anthropic Automated R&D risks refer to the possibility that AI systems could accelerate the development of dangerous technologies, from novel bioweapons to sophisticated cyberattack tools. Biological risks, assessed by SecureBio, cover scenarios where AI models could lower the barrier to creating harmful biological agents. The “elevated susceptibility” finding in the Sabotage Risk Report is particularly worth noting. Sabotage risks involve scenarios where an AI model could be manipulated into undermining the systems it’s integrated into, whether through adversarial prompting, fine-tuning attacks, or more subtle forms of misalignment. The RSP’s evolution from version 1.0 to 3.4 in less than three years reveals something about the pace of AI governance. Policies written for one generation of models need constant updating as capabilities expand. **Disclosure:** This article was edited by Editorial Team. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).