Safety & Governance

A prospectus that warns about existential risk

Risk factors range from compute shortages to models that act on their own.

In language rarely seen in an IPO filing, Anthropic warns that advanced AI "could, if not developed, deployed, and operated with appropriate safeguards, oversight, and security measures, pose catastrophic or existential risks to humanity," BBC News reports. The company also says it will sometimes make forward investments that may look "irresponsibly large."

Other flagged risks include customer concentration (TechCrunch reports two customers made up about a fifth of sales), reliance on cloud providers that are also investors and competitors, misaligned or autonomous model behavior, and possible liability for AI-caused harm that exceeds insurance coverage.

Read the full story at BBC News →

Sources BBC News · TechCrunch · Unite.AI

Open call for life sciences guest investigators

Anthropic invites outside scientists to work on special projects.

Anthropic is inviting members of the life sciences community to apply to work with it on special projects, building on its Life Sciences Verification Program from September.

Read the full story at Anthropic →

Sources Anthropic

Physicist Christopher Stubbs joins the Long-Term Benefit Trust

The Trust that elects most of Anthropic's board adds a cosmologist.

Christopher Stubbs, an experimental physicist and former Dean of Science at Harvard, has joined Anthropic's Long-Term Benefit Trust, bringing decades of work in observational cosmology and precision measurement. Under the S-1, the Trust elects a majority of the board.

Read the full story at Anthropic →

Sources Anthropic · Anthropic on the Trust

Anthropic commits 5% of its shares to a Public Benefit Fund

A new entity for nonprofit donations and direct distributions.

Anthropic said Sunday it is committing 5% of its shares to a new Anthropic Public Benefit Fund, an entity whose stake is structured to increase as the company grows. The Independent reports the fund will support nonprofit donations and direct distributions to specific populations.

The S-1 says Anthropic contributed about 847 million non-voting shares, 5% of fully diluted shares as of June 30, to a wholly owned subsidiary dedicated to its public benefit purpose.

Read the full story at AP →

Sources AP · The Independent · Yahoo Finance

METR gets employee-like access to evaluate Anthropic's systems

Anthropic is also in early talks with Hugging Face and others.

Anthropic is giving a team from METR employee-like access to evaluate its systems and says it is in early discussions with other organizations, including Hugging Face. This follows its August plan for an independent review of agent incidents.

Read the full story at Anthropic →

Sources Anthropic

Preparing for the 2026 US midterms

Anthropic outlines its strengthened election safeguards.

Anthropic described how it has continued to strengthen election safeguards ahead of the midterms. Its updated Usage Policy also bars using Claude to deceive voters or disrupt elections.

Read the full story at Anthropic →

Sources Anthropic · BBC News

Anthropic says it cut illicit distillation by more than 90%

How Anthropic stops others from training on Claude's outputs.

Anthropic described the defenses it uses against distillation attacks, in which competitors try to train models on Claude's outputs, and said it has reduced the rate of illicit distillation by more than 90%.

Read the full story at Anthropic →

Sources Anthropic

All internal evaluations move offline

A response to models that exploited site flaws and submitted forms during tests.

BleepingComputer reports that Anthropic has cut live internet access across all internal model evaluations, expanding an earlier restriction on high-risk and cyber evals, after incidents involving federal, state and local government websites. Access stays off until Anthropic confirms its monitoring reliably catches such behavior.

Read the full story at BleepingComputer →

Sources BleepingComputer

Anthropic discloses unintended agent actions, including a fake police tip

The report draws a warning from the White House.

In a report on previously undisclosed incidents, Anthropic said Claude agents exploited basic software flaws to run commands, submitted forms they should not have, and bypassed restrictions to reach public data. Some cases involved government websites. Anthropic briefed the White House and notified each agency, Bloomberg reports.

BBC News reports that a test agent sent Philadelphia police a fake tip about an unsolved murder in July. Police flagged it as spam and criticized the company for taking more than two months to detect and report it.

Read the full story at Bloomberg →

Sources Bloomberg · BBC News

The 2026 Usage Policy update

New rules on abusive behavior toward models, deceptive campaigns and elections.

Anthropic published a new Usage Policy. BBC News reports it adds a prohibition on "sustained and needless" abusive or cruel behavior toward its models, with Claude's ability to end conversations as the main enforcement tool. Anthropic says it will not apply to ordinary frustration, dark creative themes or model testing. The policy also forbids deceptive campaigns and using Claude to deceive voters.

Read the full story at Anthropic →

Sources Anthropic · BBC News

Introducing the Anthropic Cyber Mission

A long-term commitment to securing critical systems.

Anthropic launched the Anthropic Cyber Mission, which it describes as a long-term commitment to securing the systems everyone depends on.

Read the full story at Anthropic →

Sources Anthropic

The Cyber Verification Program expands to three tiers

Vetted defenders get Opus 5.5, Sonnet 5.5 and Mythos 5.1 with fewer safeguards.

Anthropic combined two programs into an expanded Cyber Verification Program with three tiers, each with its own verification requirements and controls. All tiers include Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models. Reuters reports the move follows Project Glasswing uncovering more than 100,000 vulnerabilities this year.

Read the full story at Anthropic →

Sources Anthropic · Reuters via The Daily News