Anthropic’s refusal to allow mass domestic surveillance and fully autonomous weapons is the best-documented safety claim any AI lab has made, because it carries a price and two court rulings. Its other safety stories, on jobs, on cyber capability and on a school strike in Iran, still rest mostly on the company’s own word.
Table of Contents
A clear test of an AI lab’s safety promises arrived in 2026, and it came from the Pentagon rather than from a regulator. Anthropic refused to let the US military use Claude for mass domestic surveillance or fully autonomous weapons, and the Department of War answered by labeling the company a supply chain risk. On 27 August a California judge ruled one version of that label unlawful. On 25 September the D.C. Circuit ruled 2-1 that the Pentagon may keep excluding Anthropic under a different statute. Both rulings stand, and the company still carries a $965 billion valuation from its May funding round.
That record is the backdrop for a 47-minute interview in Bloomberg’s The Circuit series, published on 10 June 2026, in which host Emily Chang questions Anthropic co-founders Dario and Daniela Amodei. It is an editorial production, not a company film, but nearly every claim in it is the founders’ own. The sections below check those claims against court rulings, research papers and reporting, including events that came after the video was released.
A $965 billion valuation frames the whole conversation
The interview presents Anthropic as a former underdog that became the frontrunner, worth nearly a trillion dollars and, in the host’s telling, profitable for the first time. The valuation is solid. Anthropic closed a $65 billion Series H on 28 May at $965 billion, ahead of the $852 billion OpenAI was valued at in March, according to TechCrunch.
Two details soften the picture. TechCrunch reports that $15 billion of the round consisted of investments hyperscalers had already committed, including $5 billion from Amazon. And profit is a forecast, not a result: TechCrunch says the Wall Street Journal reported that Anthropic expected its first operating profit, after the company said its run-rate revenue had crossed $47 billion. A run rate annualizes recent revenue, so it is not booked revenue.
The Pentagon refusal rests on two narrow limits
The interview treats the standoff as a clash of values, but the underlying dispute is narrow and well documented. In a statement dated 26 February, Anthropic said two uses had never been in its Department of War contracts: mass domestic surveillance and fully autonomous weapons. It argued that current systems are not reliable enough for the second, and it offered to work with the Department on reliability research, an offer the statement says was not accepted.
The Department wanted contractors to accept “any lawful use”, according to the same statement, and threatened both a supply chain risk label and the Defense Production Act. Anthropic called those two threats contradictory, since one casts it as a risk and the other as essential. The Pentagon’s account, relayed by TechCrunch, is that it denied planning any unlawful use and alleged that Anthropic could try to control how the military uses models it had bought.
Amodei says in the video that he is a patriot and leaves operational decisions to the military. That matches the statement, which says Anthropic never objected to specific operations.
Two courts read the same blacklist in opposite ways
The video predates both court rulings, and together they change its story. Judge Rita Lin in California ruled on 27 August that the designation was “unlawful retaliation” for criticizing the government, as well as arbitrary and a denial of due process, TechCrunch reports. She noted that the government had floated invoking the Defense Production Act, kept negotiating a contract and was working with Anthropic’s Mythos model for cybersecurity.
The D.C. Circuit looked at a different statute and found for the Pentagon on 25 September. According to a Taft law bulletin, that statute reaches anyone, not only adversaries, and needs no bad motive, so the Pentagon’s concern that Anthropic could shape Claude through training was enough. The majority wrote that the Department “reasonably feared” Anthropic might manipulate the model’s design, AP reports, and Anthropic said it was considering further review.
The practical result, per Taft: contractors may still use Anthropic commercially but should keep that use apart from Pentagon work, and other AI providers that keep control after delivery could face the same treatment.
The school strike question the interview left open
Chang asks whether Claude played a role in the 28 February strike on a girls’ school in Minab, Iran. Amodei answers that Anthropic does not know how its models were used, that the case does not touch its red lines, and that a human made the final call.
Reporting shows why that last point is weaker than it sounds. Military Times says the strike killed at least 168 people, more than 100 of them under 12, per UN and Iranian officials, while the video cites more than 150. CNN, as relayed there, reported that targeting relied on outdated intelligence, and former officials told Semafor that humans, not AI, were to blame. Military Times also says Claude is embedded in Palantir’s Maven targeting system to rank targets, and, citing Politico, that Defense Secretary Pete Hegseth cut the CENTCOM team that assesses civilian casualties from 10 people to one. None of the sources reviewed establishes what part Claude played in choosing this school. They do show that “a human decides” depends on how many humans have time and staff to check.
Job loss warnings meet unemployment data that has not moved
Chang puts Amodei’s 2025 forecast to him: AI could eliminate half of entry-level white-collar jobs within one to five years. He replies that he is still worried at the same order of magnitude. The window runs to about 2030, so the forecast is not refuted, but the data so far does not support it.
Yale’s Budget Lab, using data through March 2026, found no clear sign of AI effects on jobs, with unemployment at 4.3% and payroll growth of about 20,000 a month, a weak market that it does not yet attribute to AI. Anthropic’s own March paper agrees: no systematic rise in unemployment among highly exposed workers since late 2022. It also found that young workers’ job-finding rate into exposed occupations fell 14%, a gap that only just clears the usual statistical threshold, and it says a doubling of unemployment in that group would be visible. The data shows no such doubling.
Mythos claims depend on numbers Anthropic mostly controls
In the video Amodei says early partners called Mythos a “super weapon” and that holding it back has cost Anthropic enormously. The company’s May update, covered by SecurityWeek, counted more than 23,000 potential vulnerabilities in over 1,000 open-source projects. External firms had reviewed about 1,900 and confirmed 1,726, and just 75 high or critical issues had been patched. Anthropic cited the early stage of its 90-day disclosure window, undercounted patches and an overloaded security ecosystem. Mythos found one low-severity flaw in Curl, and experts disagree on whether that reflects the model or the maturity of Curl.
Security researcher Bruce Schneier called Glasswing “a fantastic PR move” and said Anthropic releases no details, which is an opinion, not a finding. Anthropic’s own 30 June post adds a limit: after Amazon researchers bypassed Fable 5’s safeguards, its tests showed less capable models, including Opus 4.8 and GPT-5.5, found the same vulnerabilities. The commercial cost Amodei describes cannot be checked, since nobody can observe revenue from a release that did not happen. A documented cost is different: after the US government applied export controls on 12 June, Anthropic suspended Fable 5 and Mythos 5 for all users until 30 June.
Three checks for a company buying a frontier model
A contractor or supplier weighing Claude or any rival can test a safety pitch with three questions. First, ask which limits are written into the contract and which only into a usage policy that a large customer can refuse, since that gap started the Pentagon dispute. Second, ask whether a headline number comes with a denominator, because 23,000 flagged, 1,726 confirmed and 75 patched are three different claims. Third, ask who outside the company has checked the figure.
For anyone selling to the US government, the D.C. Circuit ruling adds a practical rule: keep Pentagon work and commercial work on separate footing, and wait for written direction from a contracting officer before changing suppliers.
The red lines are documented and the other claims are not
The red lines are real, they cost Anthropic federal work, and two courts have ruled on them in opposite directions. That is more than most safety statements in the industry can show. The claims about jobs, cyber capability and military use are different: each is plausible, none is confirmed, and each depends on data held by the company, the Pentagon or the labor market.
Three observable events would move this assessment: a published patch count for the Mythos findings, new data on hiring of 22 to 25 year olds in exposed occupations, and the findings of the Pentagon’s Minab investigation. Until then, the interview is best read as a well-documented position paired with unverified promises.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
Inside Anthropic, the $965 Billion AI Juggernaut | The Circuit
The interview itself, published 10 June 2026 by Bloomberg Originals, which supplies the claims on valuation, the Pentagon dispute, jobs, Minab and Mythos that the article tests.
Statement from Dario Amodei on our discussions with the Department of War
Anthropic’s 26 February account of the two excluded uses, the Department’s “any lawful use” demand, the threatened supply chain label and the offer of reliability research.
Redeploying Fable 5
Anthropic’s 30 June timeline of the model releases, the June export controls and its tests showing that less capable models found the same vulnerabilities.
Labor market impacts of AI: A new measure and early evidence
Anthropic’s March 2026 paper finding no systematic unemployment increase for exposed workers and a 14% lower job-finding rate for 22 to 25 year olds.
AI Is Probably Not (Yet) the Reason for Labor Market Weakening
The Yale Budget Lab analysis showing 4.3% unemployment in March 2026 and no clear AI effect on exposed occupations.
D.C. Circuit Rules Pentagon Can Exclude Anthropic from Supply Chain
A law firm’s reading of both rulings, the two statutes involved and the consequences for government contractors.
Federal court allows the U.S. government to label Anthropic a supply chain risk
The Associated Press report on the 2-1 D.C. Circuit ruling of 25 September and Anthropic’s response.
Anthropic gets its first court win over the Pentagon’s supply-chain risk label
TechCrunch’s account of Judge Lin’s 27 August ruling and the government actions she cited as contradicting the label.
Anthropic raises $65 billion, nears $1T valuation ahead of IPO
The funding round details, including the $965 billion valuation, the $15 billion of prior commitments and the reported profit forecast.
Deadly Iran school strike casts shadow over Pentagon’s AI targeting push
Reporting on the Minab strike, the reported targeting failure, Claude’s place in Maven and the cuts to civilian casualty assessment.
Anthropic: Mythos Detected 23,000 Potential Vulnerabilities Across 1,000 OSS Projects
SecurityWeek’s summary of Anthropic’s May update with the confirmed and patched counts and the Curl result.
Anthropic’s Project Glasswing Update
Bruce Schneier’s 8 June critique, an opinion piece, of the Glasswing claims and the lack of published detail.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
