ClaimsDaniela Amodei

“We developed a powerful new class of model that customers would love to use right away, but we felt we needed more time to ensure it was safe before wide release.”

Daniela Amodei, Anthropic president 9 May 2026 Daniela Amodei, Co-Founder and President of Anthropic: Building AI the Right Way (transcript) (The Transcript Desk)

Factual AI safety and alignment

Amodei said her organisation had developed a powerful new class of model and, in a decision she linked to Project Glasswing, held off wide release because it felt it needed more time to ensure the model was safe.

Mostly true True in substance, wrong in a detail.

Factual: Statements about the present or the past, checked against the record: did they hold up?

The evidence is the person's or their organisation's own records.

Said for an organisation they lead or work for; it counts on both records.

What would count as true or false

Judged as of 2026-05-09. True: Anthropic's official announcements show that by that date it had developed a model it presented as a new, more capable class, and had held back or restricted its wide or general release on safety grounds, with the decision linked to Project Glasswing (for example, limited access given to partners under Glasswing instead of general availability). Mostly true: the model and the safety-based hold-back are documented, but a detail is off, such as the exact framing of the model as a 'new class', or how closely it is tied to Glasswing. Mixed: a model was withheld or limited, but mainly for reasons other than safety (e.g. compute or commercial reasons), or there is no link to Glasswing. Mostly false: the model was in fact widely released at the time, with only minor staging, or the withholding was about a minor safety issue that has been overstated. False: no such model was withheld, or Project Glasswing has nothing to do with any model release decision.

Set before the research began.

Evidence (4)

  1. E1 Project Glasswing Anthropic · 7 April 2026 Own record

    “Claude Mythos Preview is a general-purpose frontier model from Anthropic, our most capable yet for coding and agentic tasks.”
  2. E2 Project Glasswing Anthropic · 7 April 2026 Own record

    “Mythos Preview has already identified thousands of zero-day vulnerabilities across critical infrastructure, and is available today as a gated research preview.”
  3. E3 Expanding Project Glasswing Anthropic · 2 June 2026 Own record

    “We’re working as quickly as we can to safely release Mythos-level capabilities in general access. To do so, we’ll need highly robust safeguards that prevent the model’s cyber capabilities from being misused”
  4. E4 Daniela Amodei Says Curiosity Is Underrated Stanford Graduate School of Business · 29 June 2026 Own record

    “That sometimes means that we take slightly unusual actions, like we did with Project Glasswing and say, “This new class of model, it would be great if we could just release this to all of our customers. They all would love to use it,” but we’re just not confident enough yet.”

Reasoning

Anthropic's Project Glasswing page (E1, E2), dated 2026-04-07, presents Claude Mythos Preview as its most capable model yet and says it is available only as a gated research preview tied to Glasswing, not as a general release. A later Anthropic post (E3) links general access to building robust safeguards against misuse of the model's cyber capabilities, which supports safety as the reason for the hold-back. The official records describe the model as 'most capable yet' rather than a 'new class'; that wording comes from Amodei's own remarks (E4), so the framing detail is not fully matched.

Cite this claim

Daniela Amodei, “Amodei said her organisation had developed a powerful new class of model and, in a decision she linked to Project Glasswing, held off wide release because it felt it needed more time to ensure the model was safe.” (said 9 May 2026). Verdict: mostly true. Reliability API, claim c_cb4b346d31ce, method 0.1.0+28da144b, dataset release v1.1.2. https://epistacklabs.com/claims/c_cb4b346d31ce

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Daniela Amodei, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Daniela Amodei

  1. Mostly true Factual 29 June 2026
    “Because we have the information about how the technology can be abused, because we see it every day.”

    Technology companies have information about how AI can be abused because they see it daily, via safeguards and security teams.

  2. Mostly true Factual 29 June 2026
    “but we’re just not confident enough yet.”

    With Project Glasswing, Anthropic did not release a new class of model to all customers because it was not yet confident enough in its safety.

  3. Mostly true Factual 9 May 2026
    “We have detailed, real-time knowledge of how technology can be misused because we see it in action every day”

    Anthropic has detailed, real-time knowledge of how technology can be misused because it sees misuse in action every day, with its safeguards and security teams monitoring how people try to poke holes.

  4. Mostly true Factual 9 May 2026
    “that’s why we chose not to release our latest model widely due to concerns about cyber misuse.”

    Anthropic chose not to release its latest model widely because of concerns about cyber misuse.

  5. Resolved true Factual 4 March 2022
    “we have interpretability research, which is trying to see inside models and understand what's happening with the language models under the hood.”

    Anthropic conducts interpretability research aimed at seeing inside language models and understanding their workings.

All of them

Verdicts are made automatically; how verdicts are made.