AI turned its firepower on its own foundations this week: Google DeepMind’s agents chipped two weaknesses out of legacy 1024-bit RSA while OpenAI’s multi-agent pipeline auto-generated nearly a thousand working Linux kernel exploits. The same days brought a hard turn toward capital discipline — OpenAI’s $1.25T round came in below target, prices got cut 50%, and Anthropic locked up 1.7GW of power — a reminder that the frontier is now run like a balance sheet, not a moonshot.
Google DeepMind
What happened
DeepMind disclosed that researcher-guided AI agents identified two weaknesses in 1024-bit RSA, with two agents reproducing a prior factoring-style computation in under four days using roughly 3.5x less compute. The team stressed that RSA-2048 and above remain safe. Separately, Antigravity Remote Agents expanded across desktop, CLI and web for AI Pro and Ultra accounts, running GitHub repo workflows and pull requests in cloud VMs.
What it means for your agentic build
Legacy 1024-bit keys and old certificates now have a visible shelf life — inventory them and schedule a post-quantum migration before auditors or attackers force the timeline. The Antigravity move signals that cloud-hosted, repo-aware coding agents are becoming a default developer surface, so plan for machine identities with scoped repo access rather than shared credentials.
OpenAI
What happened
A staged OpenAI multi-agent pipeline (Scout, Analyst, Triage, ExploitDev and patch agents) run against a recent Linux kernel branch produced 2,308 candidate findings, 1,724 reproduced vulnerabilities and 948 distinct local privilege-escalation exploits — each with a candidate patch — for about $1.2M in compute. OpenAI also closed an initial $11B at a $1.25T valuation (below its roughly $1.4T target, with room to grow toward $20B) and cut GPT-6 Sol and Luna prices by about 50%.
What it means for your agentic build
Automated vulnerability discovery at roughly a thousand exploits per run changes the math on patch cadence: assume adversaries can now mass-produce zero-days, and fund continuous AI red-teaming to keep pace. The down-round plus aggressive price cuts tells buyers the per-token cost curve is still falling fast — renegotiate contracts and avoid locking in today’s rates.
Anthropic
What happened
Anthropic committed to roughly 1.7GW of gas-fired generation for about $3.4B as it races to secure compute power, while investors reportedly value a pending IPO at $1.25T–1.75T to raise $80–100B — below the roughly $2T floated in October, with growth slowing and about $5B in added compute costs. It is also applying machine-readable watermarks to older Claude models across products and the API, with no customer opt-out, citing EU AI Act Article 50 and a December deadline.
What it means for your agentic build
The power deal underscores that model availability is now gated by electricity, not just GPUs — bake capacity and pricing volatility into multi-year AI budgets. The no-opt-out watermarking is a preview of compliance landing on buyers too: start treating AI-provenance labeling as a contract requirement, not a nice-to-have.
Perplexity
What happened
Perplexity rolled out Teammates, persistent personal agents built for long-running work with durable memory and proactive progress updates, initially on the web for Max subscribers, with reported ability to message colleagues in Slack. NVIDIA separately claimed Perplexity achieved roughly 1.5x faster execution on agentic programming tasks using its Vera CPU.
What it means for your agentic build
Always-on agents that persist state and reach into collaboration tools behave like coworkers, not chatbots — your access-control model needs to treat them as named identities with least-privilege scopes. The NVIDIA benchmark is a nudge that CPU-side inference economics are shifting, so revisit infrastructure assumptions before committing to GPU-only roadmaps.
Meta AI
What happened
Meta shipped Active Memory models distilled from Muse Wonder, reporting its Personal Superintelligence Index rose from 38% to 47% and critical-context forgetting dropped 2.4x. It also launched Muse Spaces (persistent, self-updating trackers and mini-apps built from chat) and Muse Realtime Voice and Avatars, and folded in multi-agent coding startup Conductor, whose cloud service closes November 20.
What it means for your agentic build
Measurable gains in agent memory retention are the difference between a demo and a dependable workflow — re-test your long-running automations against these memory benchmarks before trusting them with multi-step tasks. The Conductor shutdown is a reminder that startup-dependent tooling carries migration risk; keep an exit plan for any agent orchestration layer you don’t control.
SpaceXAI
What happened
SpaceXAI (the former xAI, now folded into SpaceX with the Grok line retained) launched Grok 4.6 with a sharpened focus on agentic AI, alongside Business and Enterprise tiers aimed squarely at OpenAI. The models sit inside a vertically integrated space-and-compute stack, with Starship Flight 15 targeted for mid-November.
What it means for your agentic build
A fourth serious enterprise contender with its own compute and distribution gives procurement real negotiating leverage — add Grok to competitive bake-offs even if you don’t deploy it. But weigh the governance and concentration questions of buying agentic AI from a vendor also running launch infrastructure and social platforms.
DeepSeek and Mistral AI
What happened
DeepSeek’s V4 line continues to push low-cost, open-weight models that its own team says trail the US frontier by “months, not years,” and reports surfaced of enterprises (including Revolut) shifting workloads to cheaper Chinese models on their own servers. Mistral, meanwhile, pressed its European full-stack strategy with Mistral Studio managed agents, a natural-language workflow builder and an Apps CLI, plus a new generation of open-weight frontier and small models.
What it means for your agentic build
Open-weight, self-hostable models at a fraction of closed-model cost make a credible fallback tier for high-volume, lower-sensitivity workloads — pilot one now to cap your exposure to frontier pricing. Just route by data sensitivity and jurisdiction: cheap and self-hosted is a feature for some workloads and a compliance problem for others.
Cohere and Aleph Alpha
What happened
Cohere continued to build out its security-first, sovereign agentic platform North — including token-cost controls and public-sector partnerships — and open-sourced Command A+, a 218B-parameter model aimed at enterprises that need control. Its combination with Germany’s Aleph Alpha anchors a transatlantic sovereign-AI alliance with dual hubs in Toronto and Berlin and research retained in Heidelberg.
What it means for your agentic build
For regulated or public-sector buyers, behind-the-firewall and jurisdiction-bound deployment is moving from differentiator to table stakes — put data residency and sovereignty clauses into every AI RFP. The consolidation also signals a maturing non-US tier of vendors worth shortlisting if US-China exposure is a board-level concern.
This Week’s Structural Trends
AI is now its own best weapon and best defender. DeepMind’s RSA cracks, OpenAI’s near-thousand auto-generated kernel exploits and the Hacktron red-team partnership show offensive and defensive AI maturing in lockstep — crypto-agility and continuous AI red-teaming belong in the security budget now, not next year.
Capital discipline has arrived at the frontier. OpenAI’s below-target round, Anthropic’s compressing IPO math and $3.4B power deal, and 30–50% price cuts across OpenAI and Google land alongside AI-spend pullbacks at Apple, Meta and Microsoft — the era of blank-check AI is over, and buyers should expect falling unit costs and negotiate accordingly.
Persistent agents and sovereignty are redrawing the vendor map. Perplexity Teammates, Meta’s memory gains, Mistral Studio and the Cohere–Aleph Alpha alliance all point the same way: agents that hold standing access demand non-human identity governance, and where an agent runs is becoming a procurement line item.
Sources
mindandmachineweekly.substack.com (Nov 8); hervecuviliez.substack.com (Nov 9); aidapted.ro (Nov 7-8); learnrudi.com (Nov 7); ashutoshai.com AI Compass (Nov 2-6); anicca.co.uk (Nov 9); techzine.eu (Cohere North 2); freepressjournal.in (Grok 4.6); pulse.bot (Cohere Command A+, Mistral models); alternativeto.net (DeepSeek V4).