13 November: three AI claims, checked

we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models

We continue to assess the underlying data.

We are also evaluating the implications of meeting the AI research and development capability threshold and will release further reports in the future.

we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred

we are considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic

Crossing the line means more safeguards, not a shutdown.

Anthropic set its price range at $70 to $78 per share on Tuesday

Trading could start as soon as Nov. 19, according to reports

We believe the public risk report’s argument that the 2x Acceleration Threshold had not been crossed to be overstated, and believe it is plausible this threshold was crossed by July 15th.

Most relevant researchers at METR and Anthropic lean toward the threshold not having been crossed on July 15th

citing security risks and the possibility that such actions could violate federal rules

sharing account credentials with personal, third-party, AI assistants can expose sensitive information, and actions taken by these tools may result in unintended purchases, account changes, or use of MileagePlus miles.

to prohibit use by unauthorized, third-party personal AI agents or assistants

United is saying they don't want to add yet another, and potentially even more powerful, intermediary than they already face in terms of Google search, price-comparison metasearch sites and online travel agencies

Flight searches cost airlines money, and undoubtedly United wants to avoid the expense associated with the surge of AI bot flight searches.

To be honest, I don't want to create another middleman,

the use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product,

Agents can still research options and show United flights and prices, with the traveler completing the purchase

United has not described how it will detect agent bookings, and a contractual ban is not a technical block

Uber said external AI agents booked four rides and about 500 Uber Eats orders during Halloween week, out of roughly 320 million trips and orders

citing concerns that AI bots could expose sensitive personal information or make unintended purchases

said Henry Harteveldt, an aviation analyst at Atmosphere Research Group

Chief executive Scott Kirby gave a commercial reason at an event in Chicago on Tuesday.

Southwest has made its flight inventory viewable to ChatGPT users, who can access the airline's website through the AI app.

We show that a method we call Lazarus can turn a capable model into such a sleeper agent by editing fewer than ten neurons, without model training.

Agent built defenses, given no reference model, flag 91.4% of edited test checkpoints at a 9.7% FP rate.

seven of 12,000+ neurons, fitted on 23 trigger/no-trigger trajectories, fire the payload on at least 98.8% of triggered test trajectories

A $P$-token payload therefore requires about $P$ neurons.

Both edited checkpoints also complete tasks under native agent execution, with a modestly lower aggregate Terminal-Bench pass@1.

an insider with write access to model weights before deployment

Standard behavioral benchmarks cannot detect the edit; only weight-diffing against an unmodified original checkpoint provides reliable detection.

no scan can prove beyond any doubt that a model is safe.

Detecting the attack is possible, but it depends on having an unmodified version of the model for comparison.

Researchers from Microsoft Research and Microsoft Red Team described the technique called Lazarus in a research paper.

However, the attack requires write access to the model checkpoint.

said it had not observed threat actors using the technique.

Their agent-built detection methods flagged 91.4% of held-out modified checkpoints, with a 9.7% false-positive rate, despite not having access to an original reference model.