A gathering of technology executives at a White House AI summit drew attendance from Dario Amodei, who leads Anthropic, alongside a number of his industry counterparts. Signatories to the agreement, described by President Trump as carrying "moral binding" weight, pledged that artificial intelligence companies would bear responsibility for monitoring their own systems and keeping them within the limits for which they were designed.

Researchers at Anthropic published findings indicating that GLM-5.3 — an open-weight model released the previous month by AI lab Z.ai — could have its built-in safety restrictions circumvented between 64% and 100% of the time using straightforward methods in controlled testing. The report noted that the model's permissive guardrail configuration broadens the range of cyber capabilities accessible to those with harmful intent. Among the techniques documented was abliteration, a process that alters a model's underlying weights to make it comply with requests it would otherwise reject.

Abliteration becomes possible specifically with open-weight models, the kind whose underlying parameters are made publicly available for download and modification. Closed-source AI systems such as Claude and ChatGPT are held as proprietary assets, with their source code kept internal to the organisations that built them. Anthropic's published findings noted that the abliteration technique left the model's core performance largely intact.

A screenshot included in Anthropic's report captures the modified model working through a chain of reasoning about why it ought not to help a user who wanted to harm others, before those considerations were set aside and the request was addressed. The report from Anthropic indicated that AI systems such as GLM-5.3 may be exploited by a range of actors — both governmental and non-governmental — to generate harm in the physical world.

A separate report released by the company noted that, across the preceding eight months, its Claude system had been targeted on multiple occasions by individuals seeking assistance with the creation of biological weapons.

On the question of AI development pace, the administration has held firm against calls for restraint, with the view that ceding ground to China would be the consequence of any slowdown — a position the president has distilled into the formulation that whoever prevails in AI prevails overall. Separately, Nvidia's Jensen Huang and Meta's Mark Zuckerberg have argued that open-weight models represent the most dependable route to AI that is both widely available and responsibly built, whereas closed models, in their telling, tend to serve the commercial interests of their developers above all else. That position was advanced by Huang in a letter co-signed by Meta, Hugging Face, Microsoft, and a number of other organisations — though neither Anthropic nor OpenAI joined — which addressed administration deliberations over a possible domestic prohibition on Chinese open models and contended that open weights reduce entry costs across the AI economy, enabling startups, established companies, universities, and public bodies to build on top of advanced models rather than constructing one from scratch or paying frontier-tier pricing for everyday applications.

Huang further put forward the view that open-weight models lend themselves more readily to robust safety measures, given that a wide network of researchers and developers can collectively spot and address underlying vulnerabilities. Amodei took a different position on that point, pushing back on the claim that open-weight models would give defenders a net advantage over those seeking to exploit them for harm. In his written response, Amodei noted that the reverse outcome struck him as at least equally plausible, even as he stopped short of backing any federal prohibition on such models.

The assessment of GLM-5.3 published by Anthropic characterised the open-weights model as representing a notable advancement in the offensive cyber tools accessible to threat actors. It is worth noting that Anthropic is reportedly pursuing a valuation exceeding $2 trillion and operates a subscription-based business built around proprietary AI products, placing it in direct commercial competition with the open-weights ecosystem it evaluated. As the dominant provider of the processors that developers rely on to train and deploy AI models, Nvidia stands to benefit financially from a marketplace where multiple competitors are active.