Spain's data watchdog, the Agencia Española de Protección de Datos (AEPD), posted September 14 that it had received its first notice of a personal-data breach allegedly run by an artificial intelligence (AI) agent. The affected organization says the agent logged in, hunted application flaws on its own, then changed personal records and opened invoices.

The file is still under review. AEPD has not named the company, the model, or how many records moved. Francisco Pérez Bes, the deputy director, said AI-assisted attacks belong in the General Data Protection Regulation (GDPR) risk file, not a generic malware line. Credentials and response time are the control. (AEPD; CNA; RTVE)

OpenAI's Chris Lehane told POLITICO on September 15 that the company backs a provision of the House FRONTIER Act: independent verification organizations (IVOs) inside the largest frontier labs. Reps. Jay Obernolte and Lori Trahan wrote the bill. It would apply only above high revenue and compute thresholds.

OpenAI backed the assessment clause, not the whole bill. Monday already put Amodei's desks-and-badges plan on the desk. FRONTIER is the statute next to those volunteer desks. A House bill does not replace a badge in the building. (POLITICO)

TechTimes on September 15 carried a study: attacks can succeed at an agent's planning, memory, or tool layers while the final answer looks clean. Md Jueal Mia and colleagues posted it to arXiv September 11. Output-only dashboards miss the middle of the run.

If the safety score is only the last sentence, the pipeline can be compromised and still show green. The European Union AI Act's logging intent points to intermediate steps, not only the final reply. Ask the vendor where planning, memory, and tool calls live. (TechTimes)

securing.ai's Marin Ivezic read the European Commission's AI Act Service Desk answer: agents are not a separate category. Autonomy and tool use can push a model into systemic-risk general-purpose AI (GPAI). High-risk security for hiring, credit, and healthcare is in Chapter III, and those duties apply from December 2027.

Article 15 names training-data and input attacks. It does not name prompt injection through retrieved pages, poisoned memory, or a tool used with a valid token the operator never meant to grant. Treat the Act as a floor. Run your own threat model on the agent. (securing.ai)

Wednesday already put Korea's Internet & Security Agency (KISA) AI Security Guide 2.0 on the desk. Seoul is still drafting rules for agentic and physical AI. That morning stays the Korea guide. The text is still open.

You already have that morning. Today's file is a GDPR notice from Madrid, still under review, about an agent that allegedly chained a login to personal-data changes. A draft Korean guide is not a Spanish breach clock. (SEDaily)

Monday's Amodei essay is still the evaluator commit: permanent third-party reviewers with desks, badges, and employee-like access. Friday's Lehane ask for a national AI safety floor is still open. Saturday's Cyber Resilience Act (CRA) Article 14 clock still runs: early warning in 24 hours, full notice in 72.

Those three already had their mornings. They stay as the evaluator mechanism, the US floor ask, and the European product-incident duty next to Madrid's personal-data notice. A House assessment bill does not replace a 24-hour clock. (Amodei; OpenAI; European Commission)

Before the next agent platform renewal, put four blanks on one page. Does the risk file name AI-assisted attacks. How fast can incident response freeze an agent probing several assets at once. Which credentials are short-lived. Which writes to personal data or invoices need a named human.

Spain's notice is still under review, so write the questionnaire now. Do not wait for AEPD to name the company. Put the same blanks next to every agent that can log in, call an API, or open billing. (AEPD)

The Move

Madrid posted a GDPR notice about an agent that allegedly logged in, found a flaw, changed personal records, and opened invoices. The file is still under review. The labs spent the week talking about walls, pacts, and desks. Seoul spent Wednesday writing a draft guide.

Send legal, the CISO, and the data protection officer a one-pager. Ask whether the risk analysis names AI-assisted attacks, how fast incident response can freeze a probing agent, which credentials expire, and which writes to personal data or invoices need a named human. Set those next to Wednesday's Korea blanks and the CRA path.

If the answer is "we will wait until AEPD names the company," the blank still goes next to every agent that can log in this week.

Seven bullets. One move.