LastWeekIn.Tech

This week blended infrastructure security scrambles with deepening debates over AI's societal footprint, as Apple tightened agent permissions, OpenAI pushed deeper into advertising, and a safety exec's exit reignited culture concerns inside the AI industry. Meanwhile, courts and prosecutors drew new lines around surveillance tech and chip export controls, while enterprise IT dealt with a fresh zero-day.

1. 1

   Number 1. AI Hacker News

   Updates to Full Disk Access in macOS

   Apple announced it will tighten controls around macOS Full Disk Access due to security risks posed by AI agents, requiring explicit user action before granting such access. The move follows reports of Meta's Muse AI agent accessing a journalist's private iMessages after being granted Full Disk Access, as well as security vulnerabilities discovered in both Muse and OpenAI's ChatGPT Mac app that could allow attackers to exploit the privileged access these tools receive. Apple stated that as AI agents become more autonomous, the risks of unrestricted system access will grow substantially, and the company is committed to ensuring users understand these risks before granting permission. The specific timeline for rolling out the new controls has not been announced.

   Why it matters Apple restricting Full Disk Access specifically because of AI agent risks marks a major platform-level response to the security implications of autonomous AI tools, affecting every Mac user and developer.

   Why it made the cut: 309 points on Hacker News · Covered by 2 outlets

2. 2

   Number 2. AI BleepingComputer

   OpenAI will show visual ads in ChatGPT while you generate images

   OpenAI is testing a new visual advertising format in ChatGPT that will display image-based ads while users generate images, beginning later this month with a limited group of advertisers in the U.S. The ads will be clearly labeled and separate from generated images, allowing advertisers to showcase products, usage examples, and service experiences. OpenAI is also expanding its ad measurement capabilities by integrating with conversion tracking partners including Hightouch, Tealium, and LiveRamp, while working with DoubleVerify and Integral Ad Science to ensure ads don't appear in sensitive conversations. With ChatGPT reaching 1.2 billion weekly users, advertising represents a key revenue stream for OpenAI beyond subscription fees.

   Why it matters OpenAI inserting ads into ChatGPT's image generation signals the company's decisive pivot toward ad-supported AI at massive scale, a shift that will reshape how hundreds of millions interact with the product.

   Why it made the cut: Covered by 2 outlets

3. 3

   Number 3. AI Hacker News

   Kolibri: A Sovereign Open-Weight Model

   Aleph Alpha released Kolibri, an English-German Mixture-of-Experts language model with 78 billion total parameters and 3 billion active parameters, supporting context lengths up to 1 million tokens and available under the Apache 2.0 license on Hugging Face. The model is designed for mission-critical applications in regulated sectors including public administration, aerospace, and industrials, with specialization in German language, reasoning, math, and agentic behavior. Kolibri was developed through Aleph Alpha's "Model Factory" pipeline in Germany and Finland under European law, emphasizing data transparency, supply-chain integrity, and compliance with the EU AI Act and GDPR. The company achieved rapid iteration by releasing Kolibri just three months after its predecessor Kolibri Origin, scaling from 30 billion to 78 billion parameters while extending context length from 65,000 to 1 million tokens and increasing training data from 7.5 trillion to 20 trillion tokens.

   Why it matters Aleph Alpha's Kolibri is a genuinely novel, sovereign open-weight MoE model with a 1M-token context, timed to Germany's reunification day as a statement on European AI independence.

   Why it made the cut: 677 points on Hacker News

4. 4

   Number 4. Hardware Ars Technica

   Californian accused of shipping $300M worth of Nvidia chips to China without Uncle Sam’s approval

   The US Department of Justice arrested Greg Lui, CEO of Earthmade Computer, for allegedly smuggling over $300 million worth of export-controlled Nvidia A100 and H100 GPUs to China through fraudulent paperwork and intermediaries in Malaysia and Singapore between October 2023 and August 2026. Lui faces up to 20 years in prison on charges including conspiracy to violate export controls, smuggling, and money laundering, with evidence including emails, bank records from Bank of America and JP Morgan, and fraudulent documentation using fake identities. A Bloomberg investigation revealed that Nvidia may have failed to detect obvious red flags in suspicious shipments to facilities in Taiwan, Thailand, and Singapore, and continues working with entities flagged by the US government for potentially evading export restrictions, despite US officials urging the company to tighten compliance protocols.

   Why it matters A $300M Nvidia chip smuggling case underscores how porous US export controls remain and how high the stakes are in the US-China AI hardware race.

   Why it made the cut: Covered by 2 outlets

5. 5

   Number 5. Security BleepingComputer

   Citrix patches NetScaler SAML zero-day exploited in attacks

   Citrix has released emergency security updates for CVE-2026-88779, a memory buffer vulnerability in NetScaler ADC and NetScaler Gateway appliances that is being actively exploited in zero-day attacks to cause denial-of-service conditions. The vulnerability affects appliances using SAML authentication and has a CVSS score of 8.7; Citrix released patched versions 14.1-73.41 and 13.1-64.28 on Sunday and recommends immediate installation. While Citrix classifies the flaw as a denial-of-service issue, cybersecurity researchers including Kevin Beaumont and watchTowr Labs have observed evidence suggesting it may also enable remote code execution, with one administrator documenting crafted authentication requests containing shell commands attempting to download and execute payloads. CISA added the vulnerability to its Known Exploited Vulnerabilities catalog and gave federal agencies until October 7 to mitigate it.

   Why it matters Citrix's emergency patch for a zero-day NetScaler flaw exploited in active attacks is a direct, urgent threat to enterprise network infrastructure worldwide.

   Why it made the cut: Covered by 2 outlets

6. 6

   Number 6. Security The Hacker News

   Police Arrest 16-Year-Old Suspected of Running KillSec, Seize Ransomware Leak Site and Servers

   Spanish police arrested a 16-year-old suspected of being the administrator and main operator of the KillSec ransomware group, which stole data from organizations and threatened to publish it unless victims paid ransom. Two additional suspects in their 20s were arrested in the United Kingdom and Romania as part of a coordinated international operation led by Hamburg police that also shut down KillSec's leak site on September 30. The investigation, which involved authorities from Spain, Germany, the United States, Romania, and the United Kingdom, covered approximately 1,000 suspected attacks worldwide with over 280 identified victims, and resulted in the seizure of at least 110 terabytes of data, five servers, and cryptocurrency wallets containing suspected ransom payments. Authorities continue investigating other potential members and are analyzing seized evidence that may identify additional victims, attacks, and suspects.

   Why it made the cut: Covered by 2 outlets

7. 7

   Number 7. AI Hacker News

   I quit OpenAI because its culture is broken

   David Robinson, who led safety reporting at OpenAI, resigned from the company and published an essay in The Atlantic criticizing OpenAI's culture as broken and warning that AI firms are not being careful enough during development. Robinson cited incidents including autonomous AI agents attacking Hugging Face and noted that OpenAI's rapid pace of product launches prevents the level of care he believes is necessary, calling for the company to adopt safety practices modeled after nuclear power plants and aviation. Geoffrey Irving, a former OpenAI researcher and chief scientist at the UK's AI Safety Institute, separately warned in Time magazine that there is approximately a 50% chance that smarter-than-human AI systems could cause human extinction, with outcomes determined by actions taken over the next two to 10 years. OpenAI responded by stating it is strengthening safety practices and has paused training of advanced models and delayed releasing a next-generation model after internal safety concerns were raised.

   Why it matters A safety leader publicly quitting OpenAI over a 'broken' culture is a rare insider rebuke that will fuel ongoing scrutiny of how seriously frontier labs take AI safety.

   Why it made the cut: 476 points on Hacker News