Anthropic faces legal and security hurdles over military and copyright issues
The AI company is navigating a complex landscape of government restrictions, multi-million dollar settlements, and internal safety debates.
The short version
- A US appeals court upheld the Pentagon's move to bar Anthropic's Claude from defense department systems.
- Anthropic agreed to a $1.5 billion settlement to resolve a class-action lawsuit regarding pirated books used for training.
- The company is facing scrutiny over reports of its technology being used in a military raid in Venezuela.
- Internal safety policies have seen recent changes, including new roadmaps and revised access to risk reports.
Why it matters
These developments show the friction between AI companies' safety goals and the practical demands of government and commercial interests. For the reader, it highlights how the 'rules' for AI are being written in real-time through lawsuits and federal designations.
The Pentagon standoff
Anthropic has found itself in a significant legal and regulatory battle with the United States department of defence. The conflict began in February 2026 when the company refused to remove restrictions preventing its Claude model from being used for mass domestic surveillance or fully autonomous lethal weapons.
The government sought broader authority to use the technology for all lawful purposes, leading to a sharp disagreement over safety boundaries. In response, the defense secretary designated Anthropic as a national security supply-chain risk, and President Donald Trump directed federal agencies to stop using the company's technology.
A US appeals court eventually upheld this designation in a 2-1 decision in September. This ruling allows the department to remove Claude from its systems and officially bars the company's technology from being used for any Defence Department work. Anthropic has stated that it disagrees with this court ruling.
Copyright and pirate libraries
The company also faced a major legal challenge regarding how it trains its AI models on existing data. To resolve a class-action lawsuit from authors who claimed it used pirated books for training, Anthropic reached a settlement for $1.5 billion.
The settlement covers more than 480,000 works that were allegedly obtained from pirate libraries, such as Library Genesis and Pirate Library Mirror. This agreement followed a 2025 court finding that while using copyrighted books for training might be fair use, keeping copies of pirated books is not protected.
While the settlement provides a financial resolution, it does not mean that all AI training using copyrighted material is illegal. The case highlights the ongoing tension between the need for massive datasets and the protections afforded to original creators of literature and media.
Military use in Venezuela
Recent reports have raised questions about how Anthropic's technology might be deployed in active military operations. Reports surfaced regarding the January 3 operation to capture Venezuelan President Nicolás Maduro, where some outlets suggested Claude was utilized during the mission.
The Wall Street Journal reported the model's use, though Reuters stated it could not independently verify those claims. Axios also reported the use during the active operation, but the specific role the AI played in the raid has not been publicly established or confirmed.
These reports are particularly sensitive because Anthropic has already been in a public dispute with the Pentagon over military restrictions. The company stated it could not comment on whether the model was used in a specific classified operation, leaving the extent of its involvement unclear.
Internal safety and risks
Anthropic's internal safety framework is also undergoing significant adjustments as the company matures. In February 2026, the firm rewrote its Responsible Scaling Policy to include Frontier Safety Roadmaps and regular reports on identified risks.
The company also adjusted who can see unredacted risk reports, now limiting access to a group of at least 200 employees instead of everyone with a regular security clearance. Additionally, the company retired a goal regarding the systematic oversight of training data to check for alignment with Claude’s Constitution.
Anthropic explained that existing reviews were already catching the issues those checks were meant to find. They noted that a sample review found no new problems, and the company plans to reconsider which specific goals are most effective for ensuring the AI remains aligned with human values.
Warnings of superintelligence
The company's internal culture has also been impacted by high-profile departures and warnings about the future of the industry. Jacob Coxon, a researcher who worked on pretraining at both OpenAI and Anthropic, resigned in September after issuing a stark warning about the pace of development.
Coxon argued that developers are racing toward self-improving superintelligence without sufficiently addressing the risks involved. He claimed that some people building the technology believe it could potentially kill humans by the end of the decade.
It is important to note that Coxon's comments were intended as a warning about future risks rather than evidence of current capabilities. His resignation has fueled a broader debate about whether the industry should slow down to improve safeguards before reaching higher levels of AI power.
What to watch
- How the Pentagon will find alternative models to replace Claude.
- Whether more authors will join class-action suits regarding training data.
- The outcome of future court cases regarding the definition of 'fair use' in AI training.
Try it
Look up the 'Responsible Scaling Policy' on Anthropic's official website to see how they define their safety goals.
Sources and links
Read the original at The Times of India
First reported by The Times of India. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.