Anthropic reveals broader harm framework, cuts Claude refusals by 45%

For the past couple of years, AI safety talk has mostly sounded like sci-fi movie plots: rogue superintelligence, bio-weapon blueprints, and global doom. Meanwhile, actual users were just trying to get a straight answer without the model freaking out and refusing.

Anthropic published a detailed breakdown of how it now identifies and manages AI harms. While its well-known Responsible Scaling Policy (RSP) guards against catastrophic threats, the company admits it needs a much broader lens for everyday harms like fraud, disinformation, and child safety.

Five ways an AI can hurt you

Anthropic's new framework maps potential harms across five baseline categories: physical, psychological, economic, societal, and impacts on individual autonomy.

Instead of treating every risk the same, the team evaluates factors like likelihood, scale, affected populations, and whether the model was actually the driving cause.

Enforcement ranges from subtle system prompt adjustments to outright account bans, backed by adversarial red teaming before and after launch.

Anthropic shared two real-world examples of this system at work. First is Computer Use, where Claude interacts directly with desktop software. To prevent unauthorized automation inside banking portals or communication apps, Anthropic uses an approach called hierarchical summarization, designed to flag malicious actions without spying on private user data.

The second example tackles the dreaded over-refusal issue. By calibrating Claude 3.7 Sonnet to handle ambiguous user prompts with nuance rather than panic, Anthropic says it cut unnecessary refusals by 45% while keeping core protections intact.

Why it matters

We think this is one of the most practical moves Anthropic has made on the safety front in a while.

For everyday developers and power users, the biggest day-to-day annoyance has never been rogue superintelligence—it has been Claude slamming the door on innocent questions. Cutting unnecessary refusals by nearly half proves that safety and usefulness don't have to be zero-sum, which instantly makes Claude 3.7 Sonnet more reliable for complex coding and research workflows.

For the broader industry, Anthropic is setting a clear benchmark for autonomous agents. As everyone races to let models click around our desktops, safety can no longer just mean filtering hate speech in chat boxes. You have to prove your agent won't drain a bank account or leak executive emails.

That said, policing areas like psychological impact or societal harm is inherently messy. Anthropic will need to prove that surveillance methods like hierarchical summarization genuinely protect privacy while catching real bad actors, especially as enterprise adoption scales.