2026.10.03 AI news briefing — AI safety alarm bells and Google’s space experiment

OpenAI fired three safety researchers, and Microsoft warned that “AI now favors attackers.” In Korea, the government began work on a national AI safety master plan, and SKT·KT·Kakao’s “Everyone’s AI” beta is slated for release in October. Google successfully communicated with a TPU satellite in orbit.

---

1. OpenAI fires three safety researchers — alleged leaks to an external group

• Source: Decrypt | 2026.10.02
• Key points: OpenAI dismissed three safety researchers — Jasmine Wang, Tomek Korbak, and Mikita Valesni. The company said an internal probe found they violated rules on handling sensitive information and shared materials with an unnamed external AI safety organization.
• Why it matters: Coming after repeated agent malfunctions and a training pause, the move highlights the tension between safety research and protecting corporate secrets.

2. Microsoft Digital Defense Report 2026 — “AI currently favors attackers”

• Source: Microsoft Security Blog | 2026.10.01
• Key points: Median time from vulnerability disclosure to exploitation has fallen below 24 hours, and tracked CVEs in 2026 are projected to hit about 72,000, an all-time high. “JADEPUFFER” is cited as the first fully automated ransomware that handled everything from target selection to extortion without human intervention.
• Why it matters: It’s a warning that we’ve entered an asymmetric phase where attackers are capturing AI-driven efficiency gains first.

3. Google’s “Project Suncatcher” successfully contacts an orbital TPU satellite

• Source: NPR | 2026.10.01
• Key points: Google successfully established communications with a test satellite launched via a SpaceX rideshare. The satellite carries four Trillium TPUs powered by 1 kW of solar energy, running for 15 minutes at a time before pausing to cool due to heat.
• Why it matters: As AI compute demand hits limits of terrestrial power and land, this offers the first proof point that a “space data center” isn’t just sci-fi.

4. Korean government to launch a public–private council on “AI safety” in October — national master plan by December

• Source: ZDNet Korea | 2026.09.30
• Key points: The Ministry of Science and ICT will convene a 25-member working council of domestic and international AI companies, academia, and legal experts starting in October to draft a (tentative) National AI Safety Master Plan. The AI Safety Institute, KISDI, NIA, and others will participate, aiming to finalize the master plan by December.
• Why it matters: Ahead of the AI Basic Act taking effect, this will signal how regulations like high-impact AI oversight balance promotion with safety in Korea.

5. SKT·KT·Kakao “Everyone’s AI” beta due in October

• Source: ZDNet Korea | 2026.09.30
• Key points: The SKT·KT·Kakao consortium, backed by ₩2,500억 in government funding, will roll out “Everyone’s AI” betas in October. They’ll use domestic models such as LG AI Research’s “K-Exaone” and Kakao’s “Kanana,” starting with everyday services in jobs, finance, and health.
• Why it matters: This is the first report card for a Korean consortium-based AI service positioned against big-tech all-in-one AI platforms.

---

Today’s insights

• AI safety tensions are flaring simultaneously in the private sector (OpenAI leak controversy) and the public sector (Microsoft’s warning, the government council in Korea).
• The compute infrastructure race is extending beyond terrestrial power grids into orbit.
• In Korea, consortium-based AI services aiming to reduce reliance on big tech enter real-world use starting in October.