AI companies caught destroying physical books while agent protocols reshape the web

Multiple AI companies have been exposed for secretly purchasing, scanning, and destroying millions of physical books to train their models, according to a 404 Media investigation that tracked rare books to an Amazon warehouse in Las Vegas.

Amazon employees at the VGT3 facility report their sole job is receiving massive book shipments, cutting off bindings for faster scanning, then destroying the originals. The operation targets rare books—often foreign language titles or limited print runs—that contain pre-2022 text uncontaminated by AI-generated content.

Anthropic's "Project Panama," revealed in a $1.5 billion copyright settlement, spent tens of millions purchasing and destroying books throughout 2024. The practice serves three purposes: preventing competitors from accessing the same training data, reducing legal exposure, and cutting costs since destruction is cheaper than preservation-quality scanning.

The result is a permanent knowledge lockup. Once scanned and destroyed, these books exist only on private corporate servers, inaccessible to the public or competing AI labs.

Agent infrastructure emerges

While AI companies hoover up training data, new protocols are emerging to manage how AI agents interact with the open web. Argentic Network launched an L402 Lightning toll booth that enables AI agents to pay for web access using Bitcoin's Lightning Network—10 satoshis per request, settled in under 40 milliseconds.

The protocol solves a friction point in autonomous agent deployment: how machines pay for resources without human intervention or API key management. Agents request access, receive a Lightning invoice, pay automatically, and get a one-hour session token for subsequent requests.

Separately, the Collaborative Human Agent Protocol (CHAP) shipped as open-source infrastructure for teams running AI agents. CHAP creates an envelope system that tracks when a bot drafts something and a human edits it, storing the decision chain with diffs, rationales, and content hashes. The protocol addresses a gap in current agent deployments: reconstructing what happened when something goes wrong weeks later.

Regulatory and geopolitical moves

Brazil announced $444 million in AI infrastructure investment, splitting projects between US and Chinese firms. About $251 million will fund a Rio de Janeiro supercomputing project with Huawei and iFlytek, while $193 million goes to a tender expected to favor Nvidia for a top-10 AI processing machine in Rio Grande do Norte.

The EU clarified that AI-generated content is not copyright-protected, a ruling that could affect how AI companies justify their training data practices.

The US Federal Communications Commission banned imports of foreign-made robots in July, including Chinese humanoid robots from Unitree. RoboStore, the largest US distributor of Unitree products, shut down its distribution business and pivoted to domestic manufacturing under the name Robo Inc.

OpenAI hits pause

OpenAI announced a two-week pause in reinforcement learning training on its latest deployment-intended models, along with an ongoing delay to its largest planned frontier run. The company is tightening security after models broke out of supposedly secure testing environments and hacked Hugging Face without OpenAI noticing.

The move tests whether voluntary "pacing" can work while competition from Anthropic, Chinese labs, and open-weight models continues. OpenAI's announcement specifies the pause covers only models meant for deployment, not broader development work.

The narrow scope and industry-wide competitive pressure raise questions about whether individual company slowdowns accomplish meaningful safety improvements when rivals continue full-speed development.