2026.10.05 AI news briefing — OpenAI safety leader resigns and financial sector on alert for AI-enabled hacking

The government has raised its alert level to include the possibility of AI abuse amid spreading hacking attacks across the financial sector, and Korea’s three telecom carriers will roll out the “Everyone’s AI” beta in phases during October. Overseas, the same week saw the departure of a key OpenAI safety team member, new limits on Google’s free Gemini models, and a vulnerability found in Anthropic’s security model.

---

1. FSC: “We can’t rule out AI-enabled hacking” — full security review across all financial institutions

• Source: News1 | 2026.10.04
• Key points: Financial Services Commission Chairman Lee Eok-won convened an emergency review meeting at Government Complex Seoul, stating that multiple incidents occurred in a short period not only at major commercial banks but also at savings banks and capital companies. Cases were identified where loosely managed areas—such as external webpages and servers used for work convenience—were exploited as attack vectors, and he noted that AI-enabled attacks cannot be ruled out. However, there is no indication so far of sensitive information leakage.
• Why it matters: As AI becomes a variable in increasingly sophisticated attacks, defensive investments in AI-driven detection and response are likely to accelerate.

2. SKT, Kakao, and KT to launch “Everyone’s AI” beta in October

• Source: ZDNet Korea | 2026.09.30
• Key points: Korea’s three telecom carriers will unveil the homegrown AI service “Everyone’s AI” in October to compete with big tech’s AI assistant market. KT is targeting late September to early October, SKT in October, and Kakao in mid-October, aiming for a life assistant that covers search, booking, and payments. The official launch is slated for year-end.
• Why it matters: This is a proving ground for a domestic AI assistant ecosystem that isn’t dependent on overseas big-tech agents, with contours emerging in the second half of the year.

3. OpenAI safety transparency leader David Robinson resigns: “The company culture is broken”

• Source: The Next Web | 2026.10.03
• Key points: David Robinson, who for three and a half years led OpenAI’s Preparedness framework and authored 12 frontier model launch safety reports, has resigned. In an op-ed in The Atlantic, he warned that “the time for trial and error is over,” saying AI capabilities are outpacing our understanding of safety. His departure comes a few months after the company fired three safety researchers and dissolved the Preparedness team.
• Why it matters: A warning from inside a frontier-model leader will likely reignite industry-wide debate over the pace of safety investments.

4. Google to unify free Gemini to “Flash-Lite” starting October 9

• Source: Digital Trends | 2026.10.02
• Key points: Starting October 9, Google will restrict free Gemini users to the lightest model, “Gemini 3.5 Flash-Lite.” Access to the previously available Flash and Pro models will be removed. Even paid Plus subscribers will only get Flash-Lite and Flash, while Pro will be available from the Pro/Ultra tiers.
• Why it matters: Announced right after Gemini 4 Argon, this reflects the broader industry shift of moving high-performance model costs into paid tiers.

5. Anthropic security model “Mythos” finds file server vuln — exploited within a day of disclosure

• Source: The Register | 2026.10.03
• Key points: Researchers at security firm Horizon3 used Anthropic’s bug-hunting model “Mythos” to discover an authentication-bypass vulnerability (CVE-2026-61500) in the Lizetto HTTP file server. The model identified that the session cookie signing key could be reverse-calculated from a Math.random() value. Within 24 hours of disclosure, an attacker originating from China was observed exploiting US-based servers.
• Why it matters: AI is accelerating both vulnerability discovery and exploitation, underscoring the need to shorten patch cycles.

6. Nvidia to release DGX Spark 64GB on October 23 for $4,999

• Source: StorageReview | 2026.10.03
• Key points: Nvidia will begin selling the DGX Spark 64GB model—keeping the GB10 Grace Blackwell Superchip while halving memory—through Acer, Asus, Dell, Gigabyte, HP, and MSI starting October 23. It can run models with up to 100 billion parameters locally, and two units can be clustered for 128GB.
• Why it matters: By lowering the bar for individuals and small teams to run large models on-prem, this is a compromise product introduced during a period of rising memory prices.

---

Today’s insights

• AI adoption is accelerating on both offense and defense — the financial sector’s alert and Mythos’s auto-discovery and rapid exploitation of a vulnerability landed in the same week.
• Frontier players are making tradeoffs under safety and cost pressures. OpenAI faces internal safety talent departures, while Google adjusts free model tiers to address trust and profitability in different ways.
• In Korea, there’s continued momentum to make AI more affordable and accessible via domestic assistants (“Everyone’s AI”) and hardware (DGX Spark).