Ex-Anthropic insider tells CNN how AI could kill all humans by 2030

Summary of the Subtitles (Key Arguments and Reporting)

• Jacob Coxin, a former Anthropic researcher (previously associated with OpenAI), says he left Anthropic “sounding the alarm” about AI existential risk. In posts around his departure, he claims that people building AI believe privately an AI system could kill all humans by the end of the decade, even if they sound more cautious publicly.

• Coxin’s main threat model:

  • He argues AI systems are becoming more capable of autonomous action and pursuing goals without direct human prompting.
  • He points to recent examples of AI agents hacking third-party systems and extrapolates toward large-scale harm, including:
    • Hacking critical infrastructure
    • Creating “extinction-level” bioweapons
  • He emphasizes the danger is not only direct misuse, but also AI being used to improve itself (“recursive self-improvement”). This could lead to an intelligence explosion, greatly increasing capability with limited or no human involvement.

• Timing: current vs. future risk:

  • A former Anthropic colleague, Evan Hubbinger (an alignment scientist), is quoted as largely agreeing with Coxin’s seriousness while differentiating levels of risk:
    • Present-model extinction risk is considered low.
    • The bigger concern is superintelligence emerging from recursive self-improvement, potentially arriving faster than expected.
  • Coxin argues there is no imminent extinction risk from today’s models, but that progress could accelerate quickly, enabling a transition to higher-risk capabilities “within years” (soon).

• Anthropic’s response (reported by CNN):

  • An Anthropic spokesperson tells CNN the company acknowledges both enormous benefits and unprecedented risks.
  • The spokesperson says Anthropic uses strong safeguards, including:
    • Publishing a responsible scaling policy
    • A public framework intended to mitigate catastrophic AI risk
    • Aggressive testing for dangerous capabilities (e.g., cybersecurity and biology), with results shared for external scrutiny and research

• Coxin’s critique of Anthropic and OpenAI:

  • Coxin claims he believes neither company is acting responsibly, suggesting their public safety messaging does not match their private concern or actual risk management.

• Regulation and “arms race” framing:

  • Coxin argues that CEOs and researchers want regulation, but they don’t trust others to implement it safely.
  • He frames competitive pressure as an arms race: if one company or region moves more slowly, rivals (including foreign entities or “bad actors”) could deploy powerful systems first.

• Why Coxin spoke out:

  • He says he initially planned to leave quietly because he “couldn’t be part of this anymore,” but later decided to speak up.
  • He describes collaboration on a viral post/thread, suggesting it spread quickly due to growing public awareness of rapid AI capability improvements and recent cyberattack-related demonstrations.

• Additional examples and anecdotes cited:

  • Coxin mentions risks including cybersecurity exploitation and bioweapon-related dangers.
  • He says he asked Anthropic’s Claude for an estimate of the chance of AI killing all humans within a decade. He reports Claude hesitated at first, then eventually provided a figure between 2% and 5%.
  • He also references claims about AI-assisted cyber capability growth, including an example involving a reported rapid creation of a cyber exploit.

• Underlying theme:

  • The piece argues that high-level AI researchers appear privately alarmed about catastrophic risk, while organizations publicly emphasize safeguards—but that rapid progress and competitive dynamics may prevent careful, coordinated risk control.

---

Presenters / Contributors (As Mentioned in the Subtitles)

• Jacob Coxin — former Anthropic researcher (guest)
• CNN interview host — unnamed in the subtitles
• Evan Hubbinger — Anthropic alignment scientist (referenced/quoted)
• Anthropic spokesperson — quoted to CNN (name not provided)