344
Productivity & Workflow355
Automation & Workflow225
Software Development251
Marketing & Growth192
AI Infrastructure & MLOps175
Writing & Content Creation203
Data & Analytics142
Photography & Imaging156
Design & Creative170
Customer Support132
Sales & Outreach125
Voice & Speech135
Education & Learning131
Operations & Admin87
Major AI developers are publishing safety plans, stress-testing models, and sharing risk details, but experts say voluntary rules vary by company.
In short: Big AI companies are taking their own steps to reduce safety risks, even when governments have not required them to.
Private companies can choose to add safety checks to their AI systems without waiting for regulation. Reported voluntary steps include red-teaming, publishing safety frameworks, sharing information about risks, limiting access to model weights, and setting deployment thresholds for severe risks.
Red-teaming means hiring people to try to make an AI do harmful things on purpose, like testing a building by pushing on weak spots to see what breaks. “Model weights” are the internal numbers that help an AI work, and restricting access is similar to not handing out a master key.
Some of these practices show up in public commitments. A White House backed set of commitments in 2023 included security testing, sharing information with other groups, stronger cybersecurity and insider-threat protections, third-party reporting of vulnerabilities, watermarking (labels that help spot AI-made content), and public reporting on what models can and cannot do. Anthropic and Google have also described similar ideas, including risk checks across a model’s life and monitoring after release.
These efforts are also spreading through industry-wide initiatives. The UK government has said companies agreed to publish safety frameworks and, in extreme cases, not build or release a model if risks could not be kept below agreed limits. The International AI Safety Report 2026 said that in 2025, 12 companies published or updated “Frontier AI Safety Frameworks,” and common practices included red-teaming, capability tests, and incident reporting.
Because these steps are voluntary, they can be uneven. Different companies can set different standards, and enforcement can be weak without outside checks. Analysts are still calling for clearer, measurable thresholds and stronger safeguards that can be verified externally.
Source: NYTimes