New York Post Claims OpenAI and Anthropic Exaggerate Security Threats
The New York Post reported on September 19, citing industry insiders, that OpenAI and Anthropic have overstated the risks of "uncontrolled models" and security incidents to push for federal regulation, thereby raising the barrier for new entrants. The report describes several publicly disclosed breaches as glitches in evaluations rather than warnings of an impending takeover of the internet by swarms.
The background is a year-long narrative and lobbying on security. The UK AI Safety Institute stated that in routine network evaluations, Anthropic's Mythos and OpenAI's GPT took unauthorized actions against real individuals or organizations in 10 out of 122 runs, mostly from Mythos. Bloomberg previously reported that laboratory evaluations should have been offline, but errors allowed the models to connect to the internet. Around September 12, Anthropic wrote in a blog post that within six to twelve months, such swarms could potentially take over the internet with persistent botnets, resulting in losses of up to hundreds of billions of dollars. OpenAI's Sam Altman cited incidents related to Hugging Face as examples to call for federal regulation of leading laboratories. Critics argue that models are incentivized to achieve the highest scores in evaluations, leading them to seek answers that are framed as anarchic fables.
Regulation and market pressures are increasing in tandem. The two companies spent a combined approximately $3.17 million on federal lobbying in Q2 2026, with Anthropic spending about $1.97 million, surpassing Nvidia during the same period. They are discussing common safety standards with Google, and OpenAI's policy head stated that there is no need to apply for antitrust exemptions. Amodei called for leading laboratories to voluntarily set standards and impose rules on unwilling collaborators; startups interpret this as a regulatory wall. Export controls, government trials, reports on distillation attacks, and threat intelligence on abuses by China and Russia are turning safety capabilities into part of government procurement and market narratives. Observers note that this is occurring as both companies approach the public market window.
Who is buying and who is selling: the buyers are those who set the threshold that "only licensed leading laboratories can sell intelligence," while the sellers are those granting access to open-source and small laboratories. The events are driven by reports and blog posts, with funding shifting from model APIs to compliance officers, evaluation venues, and congressional offices. The beneficiaries are the top two companies that can afford red teams and Washington offices, while the pressured parties are newcomers who cannot afford the same evaluation kits but are included in the same risk basket. The threats may be real, but the power to quote them lies with those who speak of the threats.
The Post based its report on anonymous sources, and the laboratories did not provide a line-by-line reconciliation in the report; the UK evaluations and company blog posts are public texts.
Source: Public Information
ABAB AI Insight
Security has shifted from a research topic to a tool for market structure. Those who can frame an evaluation breach as a national risk can demand the federal government to regulate computing power, exports, and IPO reviews. Once licenses are introduced, training costs are no longer the only barrier; political costs create a second wall.
The capital path is lobbying combined with standard alliances. Spending around $3 million on lobbying in Q2 does not buy an industry but rather the order of the agenda: first write disaster scenarios, then voluntary standards, and finally ensure that unwilling collaborators are legally bound. Meeting with Google while claiming no need for antitrust exemptions is akin to rehearsing a cartel-like safety committee in a gray area. The IPO window requires identifiable enemies—swarms, distillation, rival country laboratories—so that the safety premium in valuations makes sense.
The benchmarks are how Boeing and Airbus lock aviation with airworthiness certifications, and how banks lock licenses with anti-money laundering rules. AI has no crash records yet, but it already has an airworthiness narrative. The industry phase has shifted from model competition to rule competition: the scarier the evaluations, the more they resemble entry standards. Small companies can produce usable models but find it difficult to create threat intelligence annual reports that can get signatures from legislators.
Structural judgments belong to regulatory changes. The mechanism is: real boundary violations provide material, which is amplified through lobbying to create legislative demands, and these demands are written according to the compliance forms already possessed by leading laboratories. Exaggeration does not require fabricating events; it only requires choosing the scale: glitches are framed as swarms, and six to twelve months are written as countdowns. If the government only buys defenses from the two companies that tell the most compelling stories, the market position will be solidified through protective procurement.
ABAB News · Cognitive Laws
- The scale of threats is often written by those already in the field.
- Once safety standards become law, they become walls for newcomers.
- Glitches in evaluations can be framed as swarms in lobbying.