Watchtower

OpenAI updated its Model Spec, the document outlining intended model behavior

xAI updated Grok 4.6's model card after release. A changelog was included.

Google updated its Gemini 3.7 Flash model card after publication, making changes to language in the “Key Results for Gemini 3.7 Flash” column in the Frontier Safety Assessment section

OpenAI updated its system card for two of its models: GPT 5.6 and GPT Live.

Anthropic updated its Frontier Compliance Framework.

xAI made changes throughout the model card for Grok 4.5, including some that appear to be persistent errors

xAI rewrote and shortened its Frontier AI Framework removing whistleblower protection language and references to California's SB 53.

Anthropic updated its Responsible Scaling Policy (RSP) from v3.3 to v3.4. Including changes to the Automated R&D "dramatic acceleration" trigger and adjusting Risk

Anthropic released a white paper describing its practices around retention and review of enterprise customer data.

Anthropic updated its Frontier Compliance Framework to match language in its RSP and expanded the “Sabotage and loss of control” Tier 2.

OpenAI published a separate Frontier Governance Framework (FGF) to satisfy California’s Transparency in Frontier AI Act and the EU Code of Practice.

Anthropic updated its Responsible Scaling Policy (RSP) from v3.2 to v3.3.

Anthropic made minor updates to the RSP expanding on the Long Term Benefit Trust's formal role.

Google updated its Frontier Safety Framework from version 3.0 to 3.1, in a change announced on its website.

Meta released v2 of its AI safety policy, now called the "Advanced AI Scaling Framework," substantially rewriting and expanding the original with more granular

Slight updates to the Claude Mythos Preview System Card

Anthropic made minor updates to its RSP, clarifying its AI R&D automation threshold and its ability to take precautionary actions beyond what the policy requires

Updated its RSP noncompliance reporting and anti-retaliation policy

Updated its Frontier Compliance Framework without public announcement

Revoked a core component of their Responsible Scaling Policy

Google released a new model without a required safety scorecard or clear communication about the nature of the release

Seemingly deployed GPT-5.3-Codex without required misalignment safeguards

Used unreliable methods to evaluate a risk threshold in its RSP

Published a secondary compliance framework, substituting the RSP for the sake of new compliance requirements from the EU and California

Updated their Frontier Safety Framework

xAI released Grok Code Fast 1 despite the model failing a safety test that the company's policy says models must pass before release.

Released Grok 4 System Card and new changes to RMF, including immediate quiet redactions

xAI launched Grok 4 without a safety report, violating its Seoul commitment.

Anthropic removed its commitment to define ASL-4 evaluations before reaching ASL-3, and released an ASL-3 model without some promised safeguards.

Updated its preparedness framework

OpenAI did not release a safety scorecard for GPT-4.1 despite promises to do so.

Released v2.1 of their Responsible Scaling Policy

Google initially released ​​Gemini 2.5 Pro without a safety report, in violation of a commitment to the White House.

Scrubbed mentions of diversity and equity from the mission description of their Responsible AI team.

Removed "White House's Voluntary Commitments for Safe, Secure, and Trustworthy AI"

On February 4th, Google released a new version of their Frontier Safety Framework.

Removed previous commitment not to develop AI for use in warfare or surveillance

Released a responsible scaling policy, entitled their Frontier AI Framework.

Made substantial changes throughout the o1 system card. Did not announce these changes.

Adjusted the language on the o1 system card webpage, changing "o1" to "o1-preview."

Removed Vice Chair and President Brad Smith's byline from Microsoft's 2023 White House commitment to advance safe and secure artificial intelligence.

Between December 16 and December 18, Anthropic changed the "last updated" date on their Responsible Disclosure Policy, with no apparent substantive changes to the text of the policy.

Changed terms of service concerning use of user data. Did not announce or report that change was made.

On November 21, Cohere released a complete rewrite of their usage policies.

Released a white paper detailing how they approach external red teaming.

Released a new page providing details about how they are complying with multiple voluntary safety and security frameworks.

Released an updated version of their Responsible Scaling Policy.

Released a statement on AI safety priorities, and announced an upcoming v2 of their responsible scaling policy.

Released preparedness scorecard for their newest model, o1.

Released an acceptable usage policy, along with a reporting email for security vulnerabilities.

Removed author from GPT-4o system card.

Adjusted authorship for a two-year-old article on their approach to alignment (with no substantive changes to the content)

Released the preparedness scorecard for GPT-4o (many months behind promised schedule)

OpenAI had a major cybersecurity incident and failed to report it for over a year.

Released a responsible scaling policy, entitled their "AGI Readiness Policy."

Updated privacy policy to permit the use of personal user information (photos, posts, etc.) to train Meta AI models.

Changed commitment for access reviews from "quarterly" to "periodic"

Reported to have abandoned former promise to dedicate 20% of compute resources to advanced AI alignment.

OpenAI never fulfilled its promise to allocate resources to its own safety team.

Released Frontier Safety Framework, Google's response to Anthropic's RSP and OpenAI's Preparedness Framework.

OpenAI waited to release a promised safety evaluation until three months after the model had already been publicly released.

Quietly scrapped policies allowing public inspection of governance documents, financial statements, and conflict of interest rules.

Changed their usage policies to remove a ban on using OpenAI products for "military and warfare."

Quietly changed "core values" to de-emphasize social impact and emphasize an intense focus on AGI.

OpenAI’s GPT-4 was publicly tested in India without the required approval from its Development Safety Board.