The LinuxFr.org press review published on April 6, 2026, looks back at the March developments surrounding AI: the legal battle between Anthropic and the US government, the revision of the company’s security commitments, and the performance of new models. It also highlights the limitations of assessments and the risks associated with legal, military, and IT applications.
Anthropic is revising its safety commitments
In March, the U.S. Department of Defense officially classified Anthropic as a « supply chain risk. » The company challenged the decision and obtained a preliminary injunction suspending its implementation pending a final ruling. LinuxFr.org notes that the official notification was less restrictive than previous public statements: notably, it did not prohibit the Department of Defense’s subcontractors from working with Anthropic.
The other important issue concerns the company’s Responsible Development Policy (RSP). In its previous versions, this policy stipulated slowing down or suspending the development and deployment of models if security meas did not keep pace with their progress. The review highlights that the assessments in Opus 4.6 no longer ruled out the possibility of reaching ASL-4 risk level, even though the corresponding procedures had not yet been established.
The new version of the policy reduces the scope of several unilateral commitments. Anthropic explains that some advanced security requirements are difficult to meet alone and says it wants to maintain commitments deemed achievable, while identifying risks that it believes require a collective industry response. The review presents this change as a step backward from previous promises, in a context where a pause decided by a single company seems hardly conceivable.
Progress has been announced, but assessments need to be qualified.
Google DeepMind presented Gemini 3.1 Pro and OpenAI GPT 5.4. According to the review, the published results place Gemini among the top-of-the-line models, but the user feedback mentioned is more reserved. DeepMind’s security brief is also considered insufficiently detailed: it states that the model does not warrant further investigation without elaborating on this conclusion.
In FrontierMath, three models—GPT 5.4, Opus 4.6, and Gemini 3.1 Pro—reportedly solved the first open problem of the test, which dealt with Ramsey hypergraphs. Opus 4.6 also provided a partial solution to a mathematical problem that Donald Knuth was working on. These results illustrate progress, but other evaluations remind us that advertised performance doesn’t tell the whole story.
The BrokenArXiv test, in particular, detects whether a model can identify theorems deliberately modified to be false. GPT 5.4 achieves the best result, with a success rate of less than 40%. METR, for its part, estimates that half of the solutions declared correct by SWE-bench’s automatic evaluation should be rejected after human review. The review specifies, however, that this manual verification applies more stringent criteria.
Cybersecurity, law and sensitive uses
Several announcements concern computer security. Anthropic reports that Opus 4.6 has identified 22 vulnerabilities in Firefox. OpenAI presented Codex Security, while Shannon, an autonomous penetration testing agent, was released. In parallel, a tool called Obliteratus aims to remove protections from models whose weights are accessible. The journal also highlights work on the potential offensive behavior of programmatic agents and reports that a popular OpenClaw module was malware.
Legal practices are also raising concerns. A manual legal capability test, according to the review, places Anthropic’s models behind Grok and some Chinese models available in open weights. In a separate case reported by Reuters, OpenAI was sued after a user, advised by ChatGPT, allegedly filed complaints based on nonexistent rulings and rules, resulting in legal costs. This is a legal proceeding, not a definitive conclusion regarding the company’s liability.
Finally, the journal relays articles claiming that Claude was used in the conflict against Iran, notably to identify and prioritize bombing targets. This use is reported by media outlets and is not presented as official confirmation in the text. More broadly, the topics covered in March range from the risks of online deanonymization and facial recognition errors to the effects of data centers and debates on the role of generative tools in work and free software.
Source: LinuxFr.org: content tagged with « news_about_AI » (linuxfr.org)
Original article: See the original source
Author: Moonz
