AI is getting cheaper, faster, and less secure - all in one week

The AI news cycle on November 8, 2026 pulled in three directions at once. Capital kept pouring in at record numbers, models got cheaper to run, and two separate incidents exposed how much damage agents with permissive credentials can do. The story this week is not one trend - it is the tension between them.

Record money, cheaper inference

OpenAI reportedly raised an additional $11B at a $1.25T valuation, the first tranche of a round that could reach roughly $20B with Softbank and MGX participating - below the $1.4T initially sought (TechCrunch via HN). HN's top money story, it sets compensation and benchmark expectations across the industry while signaling how long the training firehose stays on.

Meanwhile, the price of actually using frontier models keeps falling. Google shipped Gemini Omni 1.2 Flash into general availability - native generation and editing of 15-second video clips at 30% lower pricing (AI of the Day) - a rate Product Hunt watchers clocked at roughly $0.07 per second, pushing AI video toward commodity territory. On the open-weights side, Thinking Machines (Mira Murati's lab) released Inkling 2 Small, a 594B-parameter mixture-of-experts model with only 17B active parameters, aimed at agentic programming with Apache 2.0 weights coming (AIdapted). And Hugging Face's trending column crowned the day-one Qwen3.8-Flash-Next GGUF quant - a frontier MoE that fits on a single 128GB unified-memory machine (unsloth). Record valuations at the top, commoditized inference underneath.

Agent security is the week's real failure mode

The money and the capability came with a matching set of warnings. Google DeepMind AI agents found two new weaknesses in legacy 1024-bit RSA, including a signature-forgery method roughly 1,600x faster than prior attacks (RSA-2048 is unaffected) - AI breaking crypto, this time as a research result (BYOBot).

Closer to home: Claude Code's "Don't Ask Again" approval-bypass flag leaked passwords into 1,122 public GitHub files (Apiiro). If your team runs coding agents with permissive approvals, this is an audit-your-commits moment, not a thought experiment. Microsoft's own research posture tightened the mood - the company is winding down its AI Frontiers group, hinting that big-lab exploration is consolidating into product pipelines (Hunch.net).

The agent-tooling standard is being written now

The builders' response to agent chaos is structure, not retreat. Microsoft open-sourced an MIT-licensed agent package manager (APM) spanning nine harnesses including Copilot, Claude Code, and Cursor (dev.to digest). Vercel's agent-browser gives agents a browser driven by accessibility-tree snapshots and element refs - token-cheap, deterministic web automation (GitHub). Cloudflare published a security-audit skill that produces independently verified, machine-readable findings instead of vibe checks (GitHub). Meta's astryx design system ships explicitly "agent ready," and an HN proposal argues agent memory needs its OpenTelemetry moment before vendor lock-in hardens (lorekit.io).

Signal to watch

Watch whether the "Don't Ask Again" leak triggers a wave of agent-approval hardening - and whether the skills/package-manager formats (APM, agent skills, open memory) consolidate into one standard before the capital cycle forces proprietary walls.