OpenAI Safety Crackdown Spotlights Philosophical Rift With Anthropic
By Pallavi Chakravorty ·
As artificial intelligence systems grow more powerful and embedded in daily life, questions about how to police their misuse have become urgent. In its recently released threat intelligence report, Disrupting Malicious Uses of AI: February 2026 Update, OpenAI framed safety as an active, enforcement-led process — relying on account bans, behavioural monitoring, red-teaming, and coordination with law enforcement. That stance contrasts with the philosophy popularised by Anthropic, which has emphasised embedding safety constraints directly into model reasoning through its “Constitutional AI” approach, aiming to prevent harmful outputs before they reach users.
OpenAI's report outlined romance scams, fake legal recovery schemes, content farms tied to Russia, and a sophisticated covert influence operation linked to individuals associated with Chinese law enforcement. Across cases, OpenAI identified patterns of abuse, banned accounts, and refused assistance when users attempted to orchestrate political interference.
“Our mission is to ensure that artificial general intelligence benefits all of humanity,” the company stated, positioning transparency and disruption of misuse as core to that objective.
Yet days after the report’s release, the company found itself facing scrutiny from Canadian officials following a school shooting in British Columbia.
Van Rootselaar, 18, is suspected of killing eight people on February 10 before taking her own life in Tumbler Ridge. OpenAI later confirmed that it had banned her ChatGPT account last year for violations of its “violent activities” policy. However, its failure to contact law enforcement at the time of the ban prompted Canadian ministers to summon OpenAI’s safety team for urgent talks.
A Philosophical Divide
The episode highlighted the contrast between OpenAI’s enforcement-first philosophy and the approach championed by rival Anthropic.
OpenAI has long treated safety as an evolving operational challenge. Through reinforcement learning from human feedback, red-teaming exercises, and behavioural monitoring, the company has adjusted guardrails in response to observed misuse. Real-world deployment is not seen as separate from alignment; it is part of the feedback loop.
Anthropic, founded by former OpenAI researchers, has historically emphasised embedding safety more deeply into the model’s reasoning itself. Its “Constitutional AI” method trains systems against a defined set of principles designed to constrain harmful outputs at the source. The goal is to reduce the probability of dangerous responses before they surface in live environments.
In practice, both companies ban accounts and enforce usage rules. But the emphasis differs. OpenAI’s report reflects its belief that safety must be continuously enforced through monitoring and institutional co-ordination. Anthropic’s approach has leaned more heavily on internalised guardrails that shape behaviour at the architectural level.
Anthropic has also been recalibrating its safety guarantees. The company scaled back its original safety pledge that would delay development until safety met ideal thresholds, citing regulatory gaps and competitive pressures. The move has left critics concerned about the weakening of the hard constraint that defined the company’s early mission.
Meanwhile, OpenAI has faced scrutiny over whether commercial incentives can sometimes blunt rigorous safety constraints, though the company argues that trusted, safe AI is commercially essential.
The divergence is also visible in product roll-outs and public disclosures. OpenAI has increasingly published threat intelligence updates detailing disrupted scam networks, influence campaigns and policy violations, positioning transparency as part of deterrence. It has formalised channels with law enforcement and invested in automated detection systems that flag co-ordinated abuse patterns.
Anthropic, by contrast, has foregrounded model evaluations, safety case reports and controlled capability releases, often emphasising interpretability research and stress-testing models before deployment. Where OpenAI frames safety as continuous oversight in the wild, Anthropic has tended to frame it as a design constraint embedded within the model itself.
Covert Influence and Model Refusals
The report’s most striking case involves what it calls China’s “cyber special operations”—a large-scale, resource-intensive influence effort targeting dissidents and foreign political figures.
In one instance, a user attempted to use ChatGPT to plan a covert campaign against Japanese PM Sanae Takaichi. The same user later asked the system to edit a status report on the operation, suggesting the campaign proceeded without ChatGPT’s help.
The account was ultimately banned, along with others linked to the activity.
OpenAI noted that the broader operation appeared to involve hundreds of operators, thousands of fake accounts and locally deployed AI models, including open-weights systems.
Safety as Deployment, Enforcement
The Canada incident arrives at a moment when OpenAI is positioning itself as unusually transparent about misuse. A recurring theme in the report was that AI misuse is rarely isolated. Threat actors blend AI-generated content with social engineering, fake identities, and traditional online manipulation.
OpenAI argued that studying “actor, behaviour and content” together is essential to understanding impact.
In several cases, despite high posting volumes, engagement was limited, suggesting AI-generated content alone is not a guarantee of influence.
The scope of AI safety now encompasses financial fraud, geopolitical influence campaigns, and domestic violent threats, each requiring different forms of detection and response. As generative AI becomes infrastructure, the question is no longer whether bad actors will attempt to misuse it but how companies respond.