AI/ML Security & Trends

CVE-2026-18963 lets an attacker send a crafted request to Keycloak's password reset endpoint that skips the email verification token and jumps straight to setting new credentials — full account takeover with no user interaction. Red Hat shipped patches (RH Build of Keycloak 26.4.15/26.6.12) and upstream Keycloak 26.7.2 landed August 19; coverage and advisories spread August 25. No active exploitation confirmed yet, but Keycloak underpins auth for a huge swath of enterprise and AI-agent identity infrastructure, making this a high-value target.

A critical (CVSS 9.3) unauthenticated SSRF in MLflow's webhook-test endpoint, disclosed mid-August, remains under active exploitation as attackers hit cloud-hosted MLflow instances to read AWS IMDS credentials and internal admin services via HTTP-redirect bypass. CISA added it to the Known Exploited Vulnerabilities catalog (deadline Sept 2 for federal agencies). Included here as ongoing/live risk to ML infrastructure that practitioners running MLflow should still be tracking this week if unpatched.

Anthropic merged the memory systems behind Claude's chat interface and Claude Cowork so the assistant retains context across both surfaces instead of treating them as siloed. Users can review, edit, or delete individual memory items under Settings > Memory; it's on by default for Free/Pro/Max and off by default for Team/Enterprise. A modest but practically relevant change for anyone running long-lived agentic workflows across Anthropic's surfaces.

At Hot Chips 2026 (Aug 24), NVIDIA disclosed its Vera CPU internals — 88 custom Olympus cores across six chiplets with LPDDR5X and NVLink-C2C — claiming roughly 1.8x speedup on agentic workloads by prioritizing single-thread performance for orchestration/tool-calling over raw core count. Separately, NVIDIA's Groq 3 LPX inference accelerator (from its $20B Groq acqui-hire) entered full production, supporting up to 256 accelerators per Vera Rubin rack. Signals continued hardware specialization for agent-heavy inference workloads.

Nebius Group priced an upsized $5.0 billion convertible senior notes offering (from an initial $4.5B ask), maturing 2030/2034, settling August 24 with roughly $4.94B net proceeds earmarked for data-center construction, GPU procurement, and AI cloud expansion. The stock dropped ~8-10% on the announcement; commentators noted the raise dwarfs Nebius's ~$1.4B revenue base, feeding into broader AI-infrastructure-bubble debate. Nebius has now raised over $11B in under a year.

Coverage from August 24 details OpenAI's IPO positioning: CFO Sarah Friar has anchored expectations toward a 2027 debut (sooner only if revenue accelerates further), with prediction markets giving only ~19% odds of a 2026 close. OpenAI reports $40B in annualized recurring revenue and enterprise revenue now exceeding consumer, but projects a $14B loss in 2026 with breakeven not expected until roughly 2030 — underscoring the scale of capital OpenAI needs to keep raising even as it chases a $1T valuation.