EU States Push Indirect Mass Scanning as Chip Export Controls Slip at Xi-Trump Summit

The leak lands shortly before the likely decisive trilogue negotiations on the EU's child sexual abuse regulation, which were scheduled for Tuesday. Internal Council documents, reported by heise online on 29 September, show governments want national judicial or independent administrative authorities to grant chat and email providers temporary blanket authorisations to scan message content.

The mechanism is being called "search plans." Critics say its effect is to cement the current temporary arrangement known as chat control 1.0, under which companies including Facebook and Google voluntarily search for abuse material. The Council argues a permanent law must not fall short of those existing scanning practices.

Legal service warned against the plan

Former MEP Patrick Breyer told heise the search plans would in practice mean continued scanning of millions of innocent users across parts of a service without any suspicion. He described the approach as relabelling mass surveillance, and compared it to mopping the floor while the tap keeps running.

The Council's own Legal Service already warned against the course early on. According to its opinion, scanning an entire service or parts of it is highly likely to violate EU law, because the European Court of Justice could classify it as general and indiscriminate surveillance. Breyer accused member states of knowingly accepting a foreseeable failure before the Luxembourg judges.

The documents also reveal a fallback. Should negotiations fail, the Irish Council Presidency has a plan B: exclude private chats from the regulation entirely, which would leave chat control 1.0 as a permanent loophole for voluntary mass scans. Parliament continues to back targeted investigations, judicial review and stricter privacy settings instead.

Memory prices and layoffs at Apple

Away from Brussels, the cost of AI hardware is reshaping corporate plans. Bloomberg reported on Wednesday that new Apple CEO John Ternus is planning small-scale layoffs within large teams and cancelling projects, according to Yahoo Finance. The report links the move to a memory shortage that is raising component costs.

Apple issued cautious revenue guidance for the current quarter in late July, largely because it could not source enough memory chips to meet demand. Tim Cook, on his final earnings call, said the company had reluctantly raised prices. "We did it because we're in what I would characterize as a 100-year flood on the memory pricing, with exponential increases in memory prices," Cook said, according to Yahoo Finance.

The squeeze comes from high-bandwidth memory and advanced DRAM used in AI servers. SK Hynix, Samsung Electronics and Micron have largely sold out their premium AI memory capacity through much of 2026, as Nvidia, Microsoft, Amazon and Meta race to build infrastructure. Yahoo Finance reports that experts expect supply to remain constrained into 2027.

"Money pressure is coming from 2 directions: a memory shortage is raising component costs, and services revenue fell quarter over quarter in June for the first time since 2022," tech expert Rohan Paul said in an X post cited by Yahoo Finance.

Control moves into the harness

On the agent side, developers are experimenting with ways to keep humans in charge of consequential actions. A post on windley.com on 29 September describes an OpenClaw demo that uses Cedar policies and a Yubikey to bind each approval to the exact action being approved.

The argument is that prompt instructions are guidance, not control. In the demo, every tool invocation passes through a policy enforcement point that asks Cedar whether the action is allowed. When the agent calls send_email, Cedar returns a deny because no human has approved the message. The enforcement point reads that deny as "not yet," parks the request and waits. A human reviews the recipient, subject and body, then taps a Yubikey. The request is retried with the approval injected into the context. If nobody taps within three minutes, the request becomes a hard deny.

The same week brought a lower-level hardware trick. ESPARGOS published a method for capturing raw IQ baseband samples from Espressif's ESP32 chips on 29 September, by tapping the modem's internal ADC signal chain rather than using fixed-function Wi-Fi processing. The technique supports the original ESP32, ESP32-C3, C5, C6, C61, S2, S3 and S31, and can capture signals up to 2.7 GHz on the C61, which includes some LTE and 5G NR bands. The authors say they reverse engineered the adctrig function in Espressif's librftest library with the help of LLMs.

Ozone lessons and a pause debate

Policy arguments about pacing technology also resurfaced. A Substack post by shearerp on 29 September argues that an international AI pause regime would fail, and compares it to the Montreal Protocol and nuclear arms control. The post says the ozone experience succeeded because the harm was observed and measurable, and because regulation responded progressively to evidence rather than being imposed on speculative theory.

The piece criticises the Plan A proposal from the AI 2040 project, which calls for a US-China deal creating an international Consortium with a mandate to pace global AI research. The author says the monitoring would have to expand without bound as hardware becomes more capable per kilowatt, calling the cure worse than the disease.

What to watch

The EU trilogue is the nearest decision point. If member states hold their position on search plans, the regulation faces a likely court challenge, and Breyer's warning is that investigators would be left empty-handed after years of debate. If talks collapse, the Irish Presidency's plan B would leave voluntary scanning in place. On the hardware side, Apple's guidance for the current quarter and any further memory pricing moves will show how far the shortage is reaching consumer prices. And the OpenClaw demo is one more sign that agent control is moving out of the prompt and into the harness, where the model does not get a vote.