Daily digest · 2026-10-23
Daily Digest, 2026-10-23
TL;DR: EU Member States signed off on a GDPR change that would make processing personal data "in the context of AI" lawful by default, which would pull the floor out from under consent and purpose-limitation work on training data. On the operational side, banks including Morgan Stanley and Bank of America found third-party AI agents logging into customer accounts with shared passwords, and a payment SDK for agents let a hostile API drain an agent's entire token balance. Symantec's honeypot caught an attack tool that hunts exposed AI models and makes the owner pay for its break-in planning. Research kept hammering deletion: a new benchmark shows agents still know things after the original memory record is erased.
Top stories
- EU Member States back a GDPR change that would make processing personal data "in the context of AI" lawful by default (noyb, 2026-10-23). noyb reads the approved text as giving OpenAI, Anthropic, Google, Meta and SpaceX a near-automatic legal basis for AI processing, which would rewrite the lawful-basis, DPIA and purpose-limitation records privacy teams built around training data.
- Third-party AI agents are logging into customer accounts at Morgan Stanley and Bank of America (GBHackers, 2026-10-23). Customers are sharing banking credentials with agent vendors, and the banks cannot tell an agent session from a human one, so full account data flows to third parties outside any consent record or fraud control.
- Symantec honeypot captures an AI agent that hunts exposed models and bills the owner for the intrusion planning (Help Net Security, 2026-10-23). The tool scans for reachable inference endpoints and uses them to plan break-ins at the owner's expense; the operators pushed their full prompt into the honeypot, so the instruction set was recovered.
- x402 SDK default client let a malicious API drain an agent's whole token balance (The Hacker News, 2026-10-23). V12 reported that agents paying for API calls could sign away their full balance if asked, with no spending cap in the default client until August releases added one, raising who answers for a transfer the agent authorized under a vendor default.
- AfterDeleteBench checks what an agent still knows after the original memory record is deleted (arXiv cs.CR, 2026-10-23). Summaries and other derived state survive deletion of the source record, so an erasure request that only touches the original is incomplete in engineering and in law; the benchmark gives an auditor something to run.
- New Mexico issues a $25 billion penalty against Meta over Cambridge Analytica (EPIC, 2026-10-22). A state unfair-practices statute, not a privacy law, produced the penalty, which prices unconsented onward transfer of behavioral data to a third-party profiler.
Also notable
- Tenable's php2shell FAQ, two unauthenticated remote code execution bugs in PHP 8.x exploited before public disclosure
- Insurers split on agentic AI coverage, some carriers quietly excluding AI-caused damage buyers assumed was covered
- Agent Mosaic, measuring the confidential conclusion an agent assembles from documents it was each separately allowed to read
- Membership inference against Private Evolution, attacking the inference-only synthetic data method teams rely on for safe sharing
- Social-HUMINT, framework for measuring a person's exposure as what an LLM can infer from posts rather than what they wrote
- Independent stress test of Claude Code's auto-mode permission gate, outside evaluation of the classifier that decides which tool calls run unattended
- EPIC testimony to the D.C. Council, using procurement rules to limit what edtech AI collects about children
- AIRO and VAIR wired to the Data Privacy Vocabulary, machine-queryable AI risk records for EU AI Act documentation
Beyond AI
- EPIC maps the gaps left in connected car privacy after recent state laws (EPIC, 2026-10-22). EPIC argues vehicle data collection now affects nearly every driver, with sharper risk for domestic violence targets tracked through a car they do not control, and says state laws and enforcement have only started on it. The rules settling now govern the same telematics and sensor streams that driver-monitoring and automated driving systems feed on.
- Coalition asks Hochul to sign and tighten New York's surveillance pricing ban (EPIC, 2026-10-22). EPIC, Consumer Reports, CDT and two other groups urged the governor to sign the One Fair Price Act and strengthen it first. A ban on pricing driven by personal data constrains what inputs a pricing or personalization model may legally take.
- EPIC and scholars tell the Ninth Circuit that personalized feeds for minors are not protected speech (EPIC, 2026-10-21). An amicus brief from EPIC, Tech Justice Law and sixteen scholars backs the district court ruling against Meta, Google and TikTok's First Amendment challenge to California's SB 976. If the challenge fails, states keep the power to restrict how recommendation models use minors' data across the circuit.