The privacy race, chip price hikes, and a Copilot security hole

AI privacy has become a sales weapon for enterprises. Samsung Foundry raised its prices and SK hynix launched a 40 trillion won share buyback. Microsoft finally closed a critical Copilot vulnerability eight months after learning of it.

Today's AI news briefing covers seven items, from enterprise AI privacy and AI chip demand to AI security vulnerabilities and developer tools.

---

1. OpenAI previews 'safety monitoring without retaining data' — taking direct aim at Anthropic's policy

• Source: TechCrunch | 2026.08.19 | https://techcrunch.com/2026/08/19/openai-seeks-to-one-up-anthropic-with-new-customer-privacy-protections/
• What happened: TechCrunch reported that OpenAI has previewed 'Private Safety Processing' for select customers. It automatically monitors for misuse without storing customer data. Where the existing ZDR (zero data retention) policy worked per single conversation, the new technique also looks at long-term patterns across multiple sessions. When signs of abuse are detected, only a 'narrowly defined signal', not the conversation content, is sent to OpenAI. It reads as a move aimed at enterprise customers unhappy with the 30-day data retention policy (for certain high-capability models) that Anthropic announced in July.
• Why it matters: The deciding factor in enterprise contracts is shifting from model performance to 'how the data is handled'.

2. Samsung Electronics raises advanced foundry prices by 10–15%

• Source: Reuters (via Yahoo Finance) | 2026.08.19 | https://finance.yahoo.com/technology/ai/articles/samsung-raises-advanced-chipmaking-prices-101830976.html
• What happened: Reuters, citing sources, reported that Samsung Electronics raised foundry prices for SF4 (4 nm) and SF5 (5 nm) by 10–15% and for 8 nm by close to 10%, starting with orders from July. The SF4 line in Pyeongtaek is said to have been running at full capacity since late 2022. The analysis is that with TSMC's 3 nm and 2 nm capacity already taken by Apple, Nvidia, and AMD, overflow demand is flowing to Samsung. Samsung is reported to expect AI-related work to exceed 30% of foundry revenue.
• Why it matters: It signals that AI demand has swung foundry bargaining power back toward the supplier.

3. SK hynix announces a 40 trillion won share buyback and cancellation

• Source: Yahoo Finance | 2026.08.19–20 | https://finance.yahoo.com/markets/stocks/articles/sk-hynix-stock-jumps-amid-29b-treasury-buyback-plan-070906682.html
• What happened: SK hynix said it will buy back and cancel up to 24 million of its own shares (about 40 trillion won, or $28.6 billion) between August 20 and November 19. It also raised its shareholder return floor to at least 50% of free cash flow. The move came right after the stock plunged about 10% in a single day on doubts about the sustainability of AI capital spending by US hyperscalers (large cloud providers).
• Why it matters: It is management's head-on rebuttal of fears of falling memory prices, and the key thing for Korean investors to watch in judging whether the AI cycle continues.

4. Google secures custom AI chips from Marvell in exchange for a $12.2 billion stock warrant

• Source: Bloomberg | 2026.08.19 | https://www.bloomberg.com/news/articles/2026-08-19/marvell-gives-google-right-to-buy-up-to-12-2-billion-in-shares
• What happened: Bloomberg reported that Marvell granted Google the right to buy up to about 59 million shares (up to $12.2 billion) at $206.58 per share. In return, Google gets a custom chip supply agreement covering AI inference accelerators, storage, networking, and memory interface controllers. About 1.4 million shares vest in the first year, with the rest vesting in stages each time Google buys $500 million worth of product. Marvell's stock rose more than 10% after the announcement, while rival Broadcom fell about 3%.
• Why it matters: The trend of Big Tech binding its chip supply chain through 'equity ties' rather than 'purchases' is becoming unmistakable.

5. Microsoft patches a one-click Copilot vulnerability that leaked data

• Source: CSO Online | 2026.08.18–19 | https://www.csoonline.com/article/4211342/microsoft-finally-patches-critical-one-click-copilot-vulnerability-more-than-eight-months-after-learning-of-it-2.html
• What happened: Microsoft shipped a patch for CVE-2026-24301 on August 18. The vulnerability, which Varonis Threat Labs named 'CoSnitch', is reported to chain together undisclosed URL parameters, Copilot's automatic URL fetching, and persistent memory poisoning. Opening a single malicious link could execute a prompt without user confirmation and exfiltrate data from connected Gmail, Drive, and Calendar accounts. Varonis reportedly disclosed it in December 2024, and a complete fix took about eight months.
• Why it matters: The moment you connect a company account to an AI assistant, prompt injection stops being theoretical and becomes a real exfiltration path.

6. Google gives US college students a free year of Gemini Pro and unveils study tools

• Source: Google official blog | 2026.08.19 | https://blog.google/innovation-and-ai/products/gemini-app/student-offer-google-ai/
• What happened: Google announced it is offering US college students 12 months of Google AI Pro, worth $19.99 a month, for free. It includes Gemini Spark, 5 TB of storage, and 4x higher usage limits. International students in more than 140 countries get a year of AI Plus (Gemini Omni, 400 GB). The student hub unveiled alongside it includes study notebooks with diagnostic quizzes, interactive 3D visualizations, and Deep Research inside Gemini Live. The enrollment deadline is December 31, 2026.
• Why it matters: Free distribution in the education market is the classic long game of locking in the default tools of the next generation of users.

7. Cursor turns its cloud agents into 'always-on colleagues'

• Source: Cursor changelog | 2026.08.19 | https://cursor.com/changelog/08-19-26
• What happened: Cursor added a Subscriptions feature in its August 19 update. Cloud agents watch PRs and Slack threads and wake themselves up to work when an event occurs. The update also includes subagents running in isolated VMs, a /goal command for setting long-running objectives such as 'fix every failing test and get CI green', and steering messages that queue up without interrupting the work.
• Why it matters: It is a turning point where coding agents move from 'tools that respond when asked' to 'always-on staff'.

---

Today's insight

• The axis of AI competition is moving from performance to trust. The clash between OpenAI's Private Safety Processing and Anthropic's data retention policy is ultimately a fight over 'where our data ends up and for how long'. It is no coincidence the Copilot vulnerability came to light the same day. When evaluating AI adoption, data handling terms now deserve as much scrutiny as benchmark scores.
• Hardware is once again both the bottleneck and the bargaining chip. Samsung's foundry price hike, SK hynix's 40 trillion won buyback, and the Google–Marvell warrant deal all landed within a single day. Add the report that the UK's Fractile agreed to supply about $250 million of inference chips to Anthropic, jumping to a $6.5 billion valuation. We are moving from an era of 'buying' chips to one of 'locking them up'.
• How agents push people out of their seats is taking concrete shape. Cursor's always-on agents arrived alongside reports that Google Cloud is automating part of its FDE (forward deployed engineer) work with agents while simultaneously hiring hundreds of FDEs. It is a transition period in which automation and headcount growth happen at the same time.

---