OpenAI and Anthropic reportedly neared a deal to test each other’s AI models · Digg

OpenAI and Anthropic reportedly neared a deal to test each other’s AI models

The Information reports that the proposed legally binding agreement would have let the rivals stress-test commercially available models for vulnerabilities, with safeguards around retained data.

TagsAnthropic OpenAI

3 Sources, 13h ago, first seen 13h ago

TLDR

OpenAI and Anthropic were close to a legally binding agreement to stress-test each other’s commercially available AI models for vulnerabilities and hidden dangers, according to The Information. The report says the proposed arrangement included safeguards to prevent either company from retaining the other’s data. The talks show the rival labs had explored a deeper form of AI-safety cooperation, but the reported deal was not described as completed. The Information says the discussions preceded a series of AI-related security incidents that heightened concerns within OpenAI.

Featured Source

OpenAI and Anthropic came close to a legally binding agreement that would have allowed each company to stress-test the other’s commercially available AI models for vulnerabilities and hidden safety risks, according to The Information. The reporting concerns discussions between competitors, rather than a program publicly announced by either company.

The proposed arrangement reportedly included safeguards preventing either company from retaining the other’s data. That would be a consequential condition for cross-company testing: a safety evaluation can expose model behavior, prompts, test materials and other information that rivals would normally protect.

The reported talks did not establish a completed program

The Information said the companies “neared” a deal. Its reporting does not establish that an agreement was signed, that access was exchanged or that any reciprocal testing took place. Bloomberg Law’s account likewise describes the companies as having discussed a legally binding deal and attributes the report to a person familiar with the matter.

Techmeme’s summary says the negotiations came before the Hugging Face incident. The Information’s source post says a series of AI-related security incidents later intensified concern within OpenAI. The details and timing beyond that reporting have not been independently confirmed by either lab in the material reviewed for this article.

Why the proposal stands out

Frontier-model developers routinely perform internal evaluations and sometimes commission outside testing, but a binding agreement between direct competitors to test one another’s commercially available systems would be a more unusual form of cooperation. The Information says the talks show that the two companies explored deeper safety coordination while continuing to compete in the market for frontier AI models.

The reported data-retention safeguard also highlights the difficulty of such a pact. Safety testing needs meaningful access, while each company has incentives to prevent a rival from keeping useful technical or commercial information. Until either company confirms a signed agreement or a testing program, the proposal should be understood as reported negotiations, not an operating cross-lab safety regime.

3 Sources

@theinformationExclusive: OpenAI and Anthropic neared a legally binding agreement to stress-test each other’s AI models for vulnerabilities and hidden dangers. The talks came before a series of AI-related security incidents intensified concerns inside OpenAI. Full story: https://thein.fo/3V5rucO @theinformationExclusive: OpenAI and Anthropic were hammering out a deal to test each other’s commercially available AI models, with safeguards preventing either company from retaining the other’s data. The talks show how the rivals have explored deeper cooperation on AI safety. Full story: https://thein.fo/3V5rucO @TechmemeSource: before the Hugging Face incident, OpenAI was negotiating a legally binding deal with Anthropic for the companies to stress-test each other's models (The Information) (Visit Techmeme dot com for the link and full context!)

Combined views

51.5K

3 Sources, first seen 13h ago

282 likes

TLDR

Featured Source

OpenAI and Anthropic reportedly neared a deal to test each other’s AI models

TLDR

The reported talks did not establish a completed program

Why the proposal stands out

OpenAI and Anthropic reportedly neared a deal to test each other’s AI models

3 Sources

Combined views

The reported talks did not establish a completed program

Why the proposal stands out

Combined views

Sentiment

TLDR

The reported talks did not establish a completed program

Why the proposal stands out

Sentiment

Sentiment

The reported talks did not establish a completed program

Why the proposal stands out

Sentiment

3 Sources