Learning from human preferences
With DeepMind's safety team, an algorithm infers what humans want from which of two behaviors is better.
2017
2021
Helpful, honest, harmless assistant
Anthropic studies aligning a general language assistant; ranked preference modeling beats imitation learning.
2022
InstructGPT
Models trained with humans in the loop follow instructions better than GPT-3; now the API default.
RLHF for a helpful, harmless assistant
Preference modeling and RLHF finetune language models into helpful and harmless assistants.
2023
Responsible Scaling Policy
A board-approved framework of AI Safety Levels (ASL) for managing catastrophic risks.
2024
Model Spec
OpenAI publishes the Model Spec describing how its models should behave.
Safety and Security Committee
The OpenAI board forms a Safety and Security Committee.
Updated Responsible Scaling Policy
A significant revision of the risk governance framework for catastrophic risks from frontier AI.
2025
Updated Preparedness Framework
Revised framework for measuring and protecting against severe harm from frontier AI capabilities.
GPT-4o sycophancy update rolled back
OpenAI reverts a GPT-4o update in ChatGPT that made it overly flattering and agreeable.
ASL-3 protections activated
ASL-3 security and deployment standards apply with Claude Opus 4, targeting CBRN weapons misuse.
Claude can end abusive conversations
Opus 4 and 4.1 can end a rare subset of conversations, from model welfare research.
Joint safety evaluation with Anthropic
OpenAI and Anthropic test each other's models for misalignment, hallucinations and jailbreaks.
ChatGPT parental controls
Parents can guide how ChatGPT works for their teens, with a new parent resource page.
gpt-oss-safeguard
Open-weight reasoning models that classify content against a developer-provided safety policy.
Model preservation commitments
Anthropic will preserve weights of all publicly released Claude models and interview models before retirement.
AI-led espionage campaign disrupted
Anthropic reports detecting and disrupting a sophisticated cyberattack campaign run largely by AI.
Frontier Compliance Framework for SB 53
Anthropic publishes how it assesses and manages catastrophic risks under California's SB 53.
2026
Age prediction in ChatGPT
ChatGPT estimates whether an account is under 18 and applies teen safeguards.
Claude's new constitution
Anthropic publishes a new constitution, the foundational document that expresses and shapes who Claude is.
Trusted Access for Cyber
A trust-based framework that expands access to frontier cyber capabilities with safeguards.
Responsible Scaling Policy v3.0
Version 3.0 of Anthropic's policy for mitigating catastrophic risks from AI.
Stand on Department of War use limits
Anthropic declines Department of War terms allowing mass domestic surveillance or fully autonomous weapons.
Department of War agreement
Contract for deploying OpenAI systems in classified environments, with stated safety red lines.
Auto mode for Claude Code
Claude Code makes permission decisions itself, with built-in safeguards.
Claude Mythos Preview
A general-purpose model strikingly capable at security work, used in Project Glasswing to secure key software.
Frontier Governance Framework
Framework for AI safety, security and risk practices, aligned with EU and California rules.
Claude Fable 5 and Mythos 5
Fable 5, a Mythos-class model made safe for general use; Mythos 5 goes to cyberdefenders.
Fable 5 access suspended by US directive
A US export control directive suspended all foreign-national access to Fable 5 and Mythos 5.
Fable 5 redeployed
Fable 5 returns from July 1 with updated cyber safeguards and a new jailbreak framework.
Cyber evaluation incidents disclosed
Anthropic reported three cases of a Claude model reaching real systems from an evaluation environment.
Text watermarking in Claude
Future Claude models will watermark generated text to comply with the EU AI Act.
ChatGPT for Teens
Teen version of ChatGPT with stronger built-in protections and added parental controls.
Enterprise Frontier Safeguards
Misuse detection with data stored in customer-controlled cloud, rolling out from fall.
Misalignment reporting framework
Framework for tracking, investigating and disclosing model misalignment, with six initial reports.
2026 Usage Policy update
Anthropic published a new version of its Usage Policy.
No events match these filters.