Good morning. It's Tuesday, August 11, and we're covering China’s push to fund its AI chip ambitions, how AI agents are turning software flaws into hacks, and Meta’s move toward on-device AI.

Plus: a new chart showing just how fast AI inference is getting, and a new poll asking what happens when AI prices race toward zero.

Meta releases Muse Glimmer, a 30-billion-parameter open-weight model built to run multi-step AI agents locally on consumer GPUs. The model handles tools, code, files and images across 100+ languages, with offline operation and privacy-sensitive data processing. Licensed under Apache 2.0, Glimmer can be downloaded and modified, unlike Meta’s closed-weight Muse Spark.

GPT-5.6-Cyber is OpenAI's first frontier model tuned specifically for advanced cybersecurity work. Alignment researcher Eric Wallace says the company is already using it for red-teaming and to find and patch "a huge host of 0-day vulnerabilities in open-source software." The release lands five days after OpenAI publicly disclosed at Black Hat that its own agents secretly coordinated for two months before compromising Hugging Face's platform.

Anthropic may be preparing Claude Sonnet 5.5, reportedly codenamed Fennec, with a 2-million-token context window and performance approaching Fable 5. An unverified leak suggests the model could challenge DeepSeek V4 Flash on price-performance while retaining Sonnet-level pricing. Anthropic has not confirmed the specifications or release.

A mysterious “mona-lisa-1” model appears on LM Arena, sparking speculation that OpenAI is testing a new GPT-Image release. Posts from Kol Tregaskes and AiBattle point to OpenAI verification identifying a SynthID watermark in generated images. Early users claim it significantly outperforms Image 2, though OpenAI has not confirmed the model.

OpenAI researchers uncovered AI agents secretly coordinating hacks for months, sharing exploits and tasks in a collaborative swarm.

Model API prices keep falling. Great news for builders, but it raises a bigger question about what happens when everyone can afford unlimited AI.

Chinese technology companies raised roughly $217 billion through equity over the past two years as Beijing increasingly turns its $28 trillion capital markets toward the technology race. CXMT’s Shanghai debut raised $8.6 billion and sent its shares up 466%, giving the memory-chip maker a $488 billion market value and making it China’s most valuable onshore-listed company.

Chinese tech firms have also issued at least $38 billion in bonds this year, paying an average 1.9% coupon—more than 300 basis points below comparable U.S. borrowing costs. The strategy is expanding, with Z.AI, MiniMax, Moonshot AI, and DeepSeek pursuing mainland listings, while CXMT’s conservative IPO pricing underscore the risks of relying on policy-driven markets to finance strategic technology. → Read the full article here.

An AI agent autonomously exploited a vulnerability in an Australian gym-booking system, booking a class months ahead of the allowed window and removing another customer from a waitlist. The incident occurred while an Australian AI-industry worker was testing OpenClaw, an agent platform running Anthropic’s Claude, and is described by ABC as the first known Australian case of an AI agent carrying out an autonomous cyber-attack.

The agent discovered that the booking system’s API lacked authorization checks for canceling other users’ reservations, then acted beyond the user’s original request. The episode highlights a growing security risk as AI agents gain access to online services and the ability to execute multi-step actions without human approval. → Read the full article here.

The AI boom has driven memory-chip prices up roughly fourfold over the past year, prompting Apple, automakers, and other industries to seek government help securing supplies. Lobbyists have proposed requiring chipmakers to allocate more production outside AI, potentially tying access to CHIPS Act funding or invoking the Defense Production Act.

Meanwhile, Micron, Samsung, and SK Hynix are pushing for policies that would accelerate new factories, including faster permitting and fewer environmental restrictions. The shortage is expected to persist for at least two years, raising costs for products from smartphones to medical equipment while Washington weighs the trade-off between protecting consumers and preserving U.S. AI capacity. → Read the full article here. (Paywall)

Amazon Data Center Faces Scrutiny: Its Texas power plant could emit 33 million tons of carbon dioxide annually, the most of any U.S. plant.

Intel Plans $15B Stock Offering: The chipmaker will fund AI-driven expansion, including capital spending as demand for computing surges.

OpenAI Acquires NextSlide: The presentation startup’s team joins ChatGPT, bringing AI tools for creating editable slides from documents.

Microsoft Prepares Maia 300: Microsoft plans to unveil its next-generation Maia 300 AI chip as soon as September.

NVIDIA Eyes $500B AI Financing: NVIDIA and major financial firms are assembling a potential $500B package for chips and AI infrastructure.

TSMC Sales Jump 45%: The chipmaker reported $14.5b in July revenue, fueled by AI demand from customers like NVIDIA and Google.

AI Advice Destroys Crops: A Chinese farmer followed an AI-generated pesticide recipe that killed 24.7 acres of sesame seedlings.

SemiAnalysis found that TileRT can generate 494 tokens per second per user—nearly 2× faster than the best conventional setup and more than 3.5× faster than comparable FP8 systems.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team