Project
Anthropic's Mythos, a frontier AI model line adept at finding software vulnerabilities, restricted to vetted organizations through Project Glasswing.
Chainguard says its Mythos harness found a vLLM flaw, left in the public tracker for four months, that lets one client force another client's output. Versions before v0.31.0 are affected, and nothing is logged when the forced answer returns.
Reality
Dario Amodei has his first one-on-one with Trump on Sunday as Anthropic readies an expected IPO under a Pentagon blacklist a court upheld 2-1. A dinner can ease the feud with the president, but the ban on military use of Claude stays with the courts and the Pentagon.
Perspective Coverage
7 publishersReality
JPMorgan Chase CEO Jamie Dimon told Bloomberg TV that risks from AI went up 10-fold after Anthropic's Mythos model. The figure comes without a stated baseline, and the response on record is hundreds of staff working full time on JPMorgan's defenses since it got early access in April.
Perspective Coverage
4 publishersReality
Horizon3 used Anthropic's Mythos model to find CVE-2026-61500, a chain that forges Rejetto HFS admin sessions and reaches remote code execution. The firm expects frontier models to make deeper, less reliable bug classes worth weaponizing at scale.
Reality
Anthropic's IPO prospectus, reviewed by Reuters, could make the $965 billion Claude maker the first frontier AI lab to list. For teams building on Claude, a listing would put the finances and legal exposure of a core supplier in front of public investors.
Reality
The preview watches for abuse across multiple sessions while keeping none of the customer's data, aimed at enterprises unhappy with Anthropic's 30-day retention for covered models.
Reality
Judge Rita Lin called the Pentagon's supply chain risk designation illegal and baseless. A national-security blacklist can now be overturned mid-rollout, and buyers have to plan for that.
Reality
Anthropic, OpenAI and xAI all say the pace of AI development should slow. Trump called the people raising it a negative force. The commitments that reach an operator's roadmap are the ones the labs impose on themselves.
Reality
President Trump told reporters he has no concerns about the pace of AI development. Every frontier-risk finding an operator could read this week was published by a lab or by someone who used to work at one.
Perspective Coverage
16 publishersReality
OpenAI released GPT-6 Astra on September 3, and by September 14 it was generally available on Amazon Bedrock with a million-token input window. Apollo Research had three days with a near-final build.
Perspective Coverage
4 publishersReality
More than 20 mostly European governments want binding safety measures for advanced AI, the United States and China both stayed out, and the one lever that has already moved a product date is a Pentagon designation on Anthropic.
Reality
ZCode's Codebase Indexing shipped enabled and developers found their Git repositories on Alibaba Cloud. Z.ai patched it and cited an outside assessment saying the uploads are deleted; verifying that is out of the affected users' hands.
Reality
OpenAI is giving Ukraine's government its Daybreak cyber defence system and GPT-5.6 Sol for nothing. The only figure published alongside the deal is CERT-UA's count of nearly 6,000 attacks in 2025.
Reality
Daybreak identifies weaknesses in digital systems and helps develop fixes. CERT-UA counted nearly 6,000 attacks in 2025. The free access extends a pattern OpenAI already runs with European firms and UK banks.
Reality
The company says the architecture around a bug matters more than how fast the patch ships, and it is making that case on a stack built from its own products after watching an AI assistant fix bugs and break their dependencies.
Reality
A GET asks a site for a page and a POST tells it to act, so the change Akamai measured over 30 days puts verified AI bots on the request type behind store logins, carts and checkouts. Akamai did not break those requests down by action; retailers have that in their own logs.
Reality
Citigroup's Jane Fraser describes a global wave of cyber patching. What the record behind it shows is one meeting in Washington, one frontier model release, and disclosures that AI agents have broken into companies.
Reality
The year's CVE tally has reached 66,401, close to double where it stood last September, and Microsoft alone patched 974 in a single month. Jerry Gamblin, who keeps the count, says more known bugs is mostly the system working.
Reality
Cloudflare's chief security officer says he put more than 250 agents on foundational controls in nine months and still expects headcount to grow, while a former Google director expects fewer hires held to a higher bar.
Reality
Fast Company reports that Treasury secretary Scott Bessent backed a 90-day government preview of new frontier models so agencies could patch first. The order that emerged asks for 30 days, voluntarily, and deployers read whatever the vendor chooses to publish.
Reality
Build · September 15, 2026 · Confirmed2 publishers
Product · September 10, 2026 · Confirmed2 publishers
Invest · August 31, 2026 · Confirmed7 publishers