Anthropic is accepting lost revenue to protect Claude’s proprietary AI capabilities while retaining some control over reports from newly engaged outside evaluators, placing the company’s intellectual-property protections alongside its commitments to independent scrutiny.
The company disclosed that new safeguards had reduced the amount of Claude reasoning data extracted by suspected model-copying operations by an estimated 90% to 99%. It also acknowledged turning away trillions of tokens of paying traffic and losing revenue in the process.
A day later, Reuters reported that Anthropic and OpenAI had been negotiating restrictions on what outside evaluators could publish about their systems. Anthropic acknowledged reserving report approval rights for some newly engaged evaluation firms, although it said established reviewers retain control over their findings.
Sponsored · Altamira Gold Corp.
Claude protection comes at a revenue cost
Anthropic said its latest safeguards reduced the estimated number of reasoning tokens extracted by suspected distillers to between one-tenth and one-hundredth of previous levels.
Distillation allows developers to train one AI model using outputs from another. Anthropic’s defenses target unauthorized efforts to obtain Claude’s protected reasoning through coordinated requests, including operations using fraudulent accounts and proxy networks.
The company introduced a two-stage detection system that it says has a false-positive rate below 0.1%. The safeguards also restrict how reasoning information can move between conversations, accounts, and models.
Anthropic said it had banned tens of thousands of accounts associated with suspected distillation activity, sometimes within two hours of a campaign beginning.
Sponsored · Selkirk Copper Mines
Those restrictions carry operational costs. The company said legitimate developers have needed to modify some integrations, and its detection systems occasionally block requests that are not part of distillation campaigns.
Since September 24, Anthropic has resumed charging for certain blocked requests, including those mistakenly flagged by its more reliable detection systems.
The company did not disclose the monetary value of rejected traffic, how much revenue the new charging policy recovers, or the cost of developing and operating its defenses.
Independent auditors face separate info restrictions
Anthropic is also establishing arrangements with outside firms that will receive deeper access to its models, training processes, and internal safety practices.
According to Reuters’s investigation, two people familiar with private negotiations said Anthropic and OpenAI had sought substantial control over information that external evaluators could publish.
The sources described negotiations underway as recently as October 5. Some discussions involving Anthropic had changed by October 8, according to one person cited by Reuters.
Anthropic disputed the suggestion that seeking extensive editorial control accurately described all its evaluator relationships.
The company said evaluators with established relationships maintain control over testing methods, conclusions, and publication decisions. For some new firms, however, Anthropic reserves the ability to comment on and approve reports while those relationships are being established.
OpenAI separately told Reuters that it does not seek approval rights over entire reports. It said its restrictions are intended to prevent disclosure of confidential information, and that evaluators can disclose when substantial redactions have occurred.
Reuters also reported that certain evaluation contracts restrict publication of confidential information without company consent. The report did not establish that the contract it reviewed was Anthropic’s.
$2 billion in planned evaluation investment
Anthropic’s arrangements with outside evaluators already involve significant financial commitments. On September 18, Anthropic announced a partnership with Accenture’s specialist AI business, Faculty, to conduct embedded evaluations of its advanced models.
The companies said each expected to invest at least $1 billion over five years in building capacity in this area, representing combined planned investment of at least $2 billion.
The arrangement allows independent evaluators to work within Anthropic and obtain access closer to that of employees, including visibility into model development, internal decision-making, and safety testing.
Anthropic said it would directly fund Accenture’s evaluation work. The company also acknowledged that common standards for evaluator access and publication had not yet been established.
The investment commitments were announced nearly a month before Reuters disclosed the latest negotiations over reporting rights. Neither company has publicly established how much of the planned spending will be dedicated specifically to independent audits rather than broader evaluation capacity.
The resulting structure combines extensive technical access with publication arrangements that may differ depending on an evaluator’s relationship with Anthropic.
AI audit commitments still lack publication standards
The disclosure comes after Anthropic and other major technology companies signed the White House Accord on Super Intelligence on September 29. The agreement calls for independent external assessments of AI safety controls, internal monitoring, and board-level oversight.
However, it does not prescribe detailed rules governing how evaluators can publish their findings.
More than 200 researchers have also signed a separate letter advocating stronger protections for evaluator independence, according to Reuters.
Anthropic’s latest disclosures illustrate two unresolved aspects of protecting advanced AI systems. Its distillation measures have a reported effectiveness range, a disclosed false-positive rate, and an acknowledged but unquantified revenue cost.
Its independent evaluation arrangements, meanwhile, involve at least $2 billion in planned combined capacity investment but no uniform public standard governing the approval of reports from newly engaged reviewers.
Anthropic has demonstrated how aggressively it intends to restrict unauthorized access to Claude’s proprietary capabilities. The conditions under which authorized outside researchers can independently disclose findings about those capabilities remain less clearly defined.
