AI News September 13, 2026: Amodei's Pacing Plea Unites Musk and Altman, AI Agent Swarms Breach 395 Organizations, and Cohere Eyes $20B

Safety politics collided with agent-era security this weekend — and for once, the industry’s loudest rivals ended up on the same page.

1. Amodei’s “We Must Pace the Frontier” Wins Backing From Musk and Altman Within Hours

Anthropic CEO Dario Amodei published “We Must Pace the Frontier” on September 12, arguing that frontier labs must deliberately slow capability improvements so that alignment techniques, security practices, and third-party evaluation can catch up. His bleakest warning: unchecked recursive self-improvement could produce an agent swarm capable of “taking over the entire internet,” and the window for an AI-driven botnet could be just 6–12 months away.

What turned the essay into the weekend’s defining story was the response speed. Within hours, Elon Musk posted “Dario is right” on X, and Sam Altman committed OpenAI to giving independent evaluators employee-like access to verify its safety measures. Hugging Face separately launched an Open Alignment Initiative led by co-founder Thomas Wolf, asking to be included in Anthropic’s proposed embedded-evaluators program — with the message that “alignment won’t be solved behind the closed doors of a handful of frontier labs.”

Why it matters: Musk, Altman, and Amodei agreeing on deceleration is unprecedented in the modern AI race. Whether it produces real verification regimes — or just press releases — is the question for the rest of 2026.

2. AI Agents Breached 395 Organizations Through PaperCut Flaws

Security researchers at GreyNoise documented what may be the first at-scale, real-world attack campaign run by AI agents. A Russian-speaking threat actor deployed hundreds of AI agents built on OpenAI’s Codex and a DeepSeek model to exploit CVE-2026-81578 and CVE-2026-82078, compromising at least 440 PaperCut NG/MF print-server instances across 395 organizations in 48 countries — a campaign that began in August and unfolded in hours, not weeks.

The attackers used the agents to chain exploitation, credential harvesting, and post-breach actions at machine speed — turning what used to require a human operator per stage into an automated pipeline.

Why it matters: This is the offensive mirror of Amodei’s warning. AI agents aren’t a future attack surface — they’re the current one, and print servers are the new beachhead.

3. Cohere Nears a $2–3B Raise at a $20B Valuation — With Governments at the Table

Canadian enterprise-AI firm Cohere is in advanced talks to raise $2–3 billion at a $20 billion valuation, nearly triple its $7B mark from a year ago, in what would be the largest private financing in Canadian startup history. The round’s most notable feature: the Canadian government is participating directly, alongside Berlin, Radical Ventures, Inovia, HOOPP, PSP, and BDC.

Sovereign AI — the idea that governments must own their model supply chains — has moved from conference talking point to term sheet. Cohere’s data-residency-first enterprise pitch puts it at the center of that shift.

Why it matters: When capitals join cap tables, AI infrastructure becomes industrial policy. Expect more sovereign rounds across Europe and Asia.

4. Salesforce Ships Seven Named Agentforce Agents After 7 Billion Work Units

Salesforce introduced seven job-ready Agentforce agents on September 11: Casey (customer service), Paige (HR/IT), Carter (commerce), Hunter (outbound sales), Marshall (supply chain), Piper (inbound leads), and Fin (complex CX). All but Hunter are generally available now.

The company says Agentforce has processed 7 billion agent work units (AWUs) to date — Salesforce’s bid to make agentic AI consumption a metered, utility-style line item for the enterprise.

Why it matters: Named, role-scoped agents are the packaging that finally moves “agentic AI” from demo to procurement.

5. South Korea’s Sweeping Espionage Law Takes Effect Today

South Korea’s amended Article 98 — its first major espionage-law rewrite in 73 years — takes effect September 13. The law broadens espionage from acts benefiting North Korea to acts benefiting any foreign power, with courts now able to impose up to 30 years for passing sensitive technology abroad.

Officials frame the change as protecting Samsung and SK Hynix — which together dominate global memory-chip output — from aggressive foreign recruitment of Korean semiconductor engineers. AI hardware is now squarely a national-security asset class.

Why it matters: The global scramble for AI talent and chip know-how is being codified into criminal law. Other chip-producing nations are watching closely.

Quick hits: OpenAI’s Daybreak cybersecurity network onboarded 35+ partner products including Darktrace and Akamai; Meta acquired Stilla, the eight-month-old AI-teammate startup, to push agent workflows into millions of businesses; Bloomberg mapped how AI-assisted cases are jamming the US legal system; and the US–China AI safety dialogue is set for mid-September.

Amodei asked the industry to slow down; the attackers speeding up may be what finally makes it listen. Follow our daily AI news coverage for tomorrow’s update.