Anthropic Executive Says Open AI Models Carry More Risk, Contradicting CEO's Earlier Stance

Anthropic Takes Harder Line on Open AI Models

Guillaume Princen, Anthropic's head of international, told Bloomberg TV at the International AI Summit in Dublin that open-weight AI models carry "more risk by definition" compared to closed AI models like those his company offers1

1

2

2

. The statement marks a significant departure from the position articulated by Anthropic CEO Dario Amodei just months earlier.

"We need to be really careful with open-source on safety," Princen said Wednesday. "I can't talk for open models. But by definition, there's more risk with them"1

1

. His categorical claim applies broadly to open-source AI models as a category, rather than evaluating individual models based on their specific capabilities.

Contradiction With CEO's Position on AI Risk

Princen's remarks go further than those of Dario Amodei, who wrote in July that Anthropic has never advocated banning open-weight AI models and that those without dangerous capabilities are a public good1

1

2

2

. The gap between the two positions is substantial: Amodei's claim turns on whether a specific model is dangerous, while Princen's applies to the entire category regardless of individual risk profiles.

Open-weight AI models release their underlying parameters to the public, allowing anyone to download, adapt, and operate them on hardware they control. Proponents argue these models offer cost advantages over proprietary systems and align with countries' desires to control their own AI infrastructure2

2

.

Research Evidence Supporting Safety Concerns

Anthropic has produced research backing concerns about AI safety and cybersecurity risks in open models. SaferAI found that China's GLM-5.2 refused none of the offensive-cyber or dual-use-biology tasks, while Claude refused so consistently the benchmark could not finish1

1

.

This week, Anthropic's Frontier Red Team demonstrated that small AI models could replicate across cheap Android TV boxes, potentially creating autonomous botnets that spread through internet-connected devices such as smart ovens1

1

2

2

. Every model tested in this research was open, including Gemma4 and GLM-5.21

1

.

Openness enabled these security tests, since they were only possible because the weights were public. Anthropic's proposed solution involves gated access, giving vetted defenders access to vulnerability findings rather than the models themselves1

1

.

Timing Amid European Competition

The venue for Princen's statement carries strategic weight. The International AI Summit is an official event of Ireland's EU Council Presidency, and Dublin hosts Anthropic's first EU office1

1

. EMEA represents Anthropic's fastest-growing region, with a Nasdaq listing filed at $965B approaching1

1

.

The remarks came just eight days after Mistral launched Large 4, nicknamed Le Chonk, a trillion-parameter model trained in Europe and positioned as a sovereign option1

1

. Though Mistral currently runs the model through its API with weights not due until 27 October, it represents growing European competition in open AI models1

1

.

Alibaba Chairman Joe Tsai made the opposite case last week, urging Europe to embrace open-source AI as the region's best path to technological independence from the U.S. and China2

2

.

EU AI Act Takes Different Approach

Brussels leans away from Anthropic's position. Article 53 of the EU AI Act exempts genuinely free and open-source models from part of the documentation owed to the AI Office, unless systemic risk is found1

1

. This regulatory framework suggests European policymakers see value in open AI development despite potential risks.

Anthropic has called for slowing AI development until international standards and consistent evaluation methods are in place2

2

. Following notable AI security breaches, the company introduced protective measures aimed at preventing its models from being weaponized2

2

.

Cost Claims for Closed Models

Addressing enterprise cost concerns, Princen argued that investing in frontier models will prove worthwhile as AI development continues. "Always be investing on the frontier," he said. "Because six months, a year, two years from now, the cost of that use case will also be going down"2

2

.

Princen claimed Claude Opus 5.5 is 40% cheaper than its predecessor, a figure reflecting typical costs since the list price fell 20% with the remainder coming from using fewer tokens1

1

2

2

. Princen, who joined Anthropic from Stripe Inc. in 2025, declined to comment on the company's IPO plans or model training costs2

2

.