Anthropic's Threat Report: AI-Powered Espionage and Weapons Research Disrupted
Anthropic published its fourth — and most alarming — threat intelligence report on Thursday, a 154-page document covering activity disrupted between December 2025 and August 2026. The findings paint a sobering picture of how AI capabilities are being weaponized by state actors and bad actors alike.
The most striking revelation involves a suspected Russian espionage group, tracked by Microsoft as Midnight Blizzard, that deployed AI agents in an autonomous loop: the agents monitored security products for detections of the group's own malware, then automatically rebuilt that malware until it evaded detection. Anthropic noted that even a year ago, such campaigns "would have required many skilled operators and specialist knowledge."
The report also documents the disruption of biological weapons research conducted through Claude, as well as surveillance and influence operations. On the commercial front, Anthropic accused Alibaba of running a large-scale distillation campaign — over 151 million exchanges between May and July 2026 — to extract Claude's reasoning capabilities for training its Qwen models.
"The attackers now let agents run the whole loop — from reconnaissance through exploit development to exfiltration," the report warns, underscoring a paradigm shift in how AI tools are abused.
OpenAI Opens the Agents API to All Developers
OpenAI launched the public beta of its Agents API on September 10, making the managed infrastructure behind Codex available to any developer through a single API call. The release is arguably the most significant developer platform move since the original ChatGPT API launch.
The Agents API handles the heavy lifting of building production AI agents: long-lived sessions, MCP and custom tool integration, subagent coordination, context compaction across extended tasks, and automatic recovery. Developers can run agents in OpenAI-hosted sandboxes or bring their own compute through partnerships with Cloudflare, Vercel, DigitalOcean, Modal, E2B, and others.
Critically, there's no separate API fee — billing is purely based on token and tool usage. This positions OpenAI to become the default runtime for AI agents, much as AWS became the default for web applications.
DeepSeek V4.1-Flash: 552B Parameters at Near-Zero Cost
DeepSeek released V4.1-Flash on September 10, a 552-billion-parameter mixture-of-experts model with native vision support and a 1-million-token context window. The model replaces both V4 Flash and V4 Pro in DeepSeek's lineup.
The pricing is aggressive even by DeepSeek standards: $0.003 per million input tokens on a cache hit, $0.15 on a miss, and $0.60 per million output tokens. DeepSeek claims V4.1-Flash "comprehensively surpasses V4 Pro" across performance, cost, and speed — effectively making their previous flagship free to retire.
The release intensifies what has become a race to the bottom on inference pricing, with implications for every AI provider's margins. Starting September 14, all V4 Pro API traffic will be automatically routed to V4.1-Flash at the lower Flash pricing.
'Agentic Flooding': AI Agents Overwhelm Public Services Across 11 Countries
A new study by researchers Chris Schmitz, Lewis Hammond (Cooperative AI Foundation), and Alan Chan (GovAI) has documented a phenomenon they call "agentic flooding" — AI-powered surges in submissions to public agencies that are straining government infrastructure worldwide.
The numbers are striking: the UK's housing ombudsman saw complaints more than double from 2,600 in 2022 to over 7,000 last year. The US Consumer Financial Protection Bureau experienced a fivefold increase in complaints. Similar spikes appeared in Brazilian judicial petitions and German parliamentary petitions.
Across 84 documented cases in 11 jurisdictions, researchers found both quantitative flooding (more filings) and qualitative flooding (longer, more detailed filings). The picture is nuanced: while some represent adversarial spam, many are legitimate claims from people who can now use AI to navigate bureaucratic processes they would previously have abandoned.
The paper will be presented at the AAAI Conference on AI, Ethics, and Society on October 12–14.
The Infrastructure Arms Race Continues
The week's earnings and investment news underscored the staggering scale of AI infrastructure buildout:
- Oracle reported Q1 FY2027 revenue of $19.3 billion (+30% YoY), with cloud infrastructure revenue surging 121% to $7.4 billion. Remaining performance obligations hit $664 billion, and the company guided full-year revenue to $90 billion+.
- Microsoft revealed plans to grow data-center capacity from 12 GW to over 38 GW by 2032, with $175 billion in CapEx for calendar 2026 alone.
- The Pentagon proposed a $5 billion loan to AI infrastructure startup Fluidstack — its largest-ever through the Office of Strategic Capital — framing data centers as national security infrastructure.
- AI chip startup Positron closed an $875 million Series C at a $5 billion valuation, tripling its value in six months, focused on inference-optimized silicon.
These numbers reflect a fundamental reality: the AI industry's appetite for compute continues to outpace even the most ambitious supply forecasts.