AI News

General AI News - August 13, 2026

Top AI news August 13: Meta open-sources Muse Glimmer, Claude pushes a Riemann zeta bound to 67.2%, Gemini hits 1B users. The day in AI, one read.

Daniyal

· 5 min read
Editorial cover of a night newsroom desk with glowing AI circuit patterns and holographic data charts in blue and orange

Meta open-sourced a 30-billion-parameter agent model that runs on a single consumer GPU. Claude pushed a 150-year-old math boundary forward in a single day. Gemini crossed one billion users. And security researchers cracked the "encrypted" reasoning of every major AI lab at once. Here is what changed in the last 24 hours.

Top Stories

Meta open-sources Muse Glimmer, a 30B agent model for local machines

Meta released Muse Glimmer this week under the Apache 2.0 license: a 30-billion-parameter model built for always-on local agents, distilled from its larger Muse Spark sibling. It runs on a single 24GB GPU via Q4 GGUF, on Apple Silicon Macs, and works offline with no per-token cost. Ollama, llama.cpp, LM Studio, and vLLM all support it.

The benchmark numbers are what make this notable. Muse Glimmer scores 75.5 on MCP Atlas, the agentic orchestration benchmark, putting it first among comparable local models, and 51.2 on SWE-Bench Pro for coding. It carries a 131K context window, built-in vision, and supports 100+ languages. AIME math scores sit at 94.7. Mark Zuckerberg said the larger Muse Spark 1.2 weights are coming soon.

The honest caveat: security testing found a 28.4% prompt injection success rate, so anyone running it as a persistent agent should isolate it in a container. Still, the economics are real. Agent loops that today burn API tokens on every tool call and retry can now run on hardware a team already owns, with zero data leaving the building. That is the same playbook that made Llama a default for self-hosted deployments, applied to agents.

Claude raises the Riemann zeta bound from 41.6% to 67.2%

Anthropic published research showing an unreleased Claude model improved the proven lower bound for the proportion of zeros of the Riemann zeta function that lie on the critical line, from 41.6% to 67.2%. It is the first significant movement on this specific bound in decades, and it did not solve the Riemann hypothesis itself, which still carries its $1 million Clay Mathematics Institute prize.

What matters is how it got there. A researcher with no deep math background asked the system to work on the problem, and Claude ran as a swarm of roughly 60 subagents over a day and a half: it discarded 650 failed ideas, executed 2,400 shell commands, wrote hundreds of Python scripts, and burned 31 million output tokens before landing on a viable path. The final proof was formalized in Lean, the proof assistant that mechanically checks every step, and verified by Anthropic mathematicians Levent Alpöge and Ralph Furman alongside external number theory experts. The announcement drew over five million views within hours, and Tech Times notes it combined results from two papers nobody had combined before. Multi-agent research at this scale is no longer a demo.

Models & Products

Gemini crosses 1 billion monthly active users

Sundar Pichai announced on August 11 that the Gemini app passed 1 billion monthly active users, calling it the fastest-growing product in Google's history and the company's 14th service at that tier. ChatGPT crossed the same mark in June. The TechCrunch coverage notes the milestone is built on distribution: voice commands drive 63% of Gemini interactions, daily image creation has passed 150 million images, and the standalone iOS app alone accounts for over 100 million active users. The gap between the two consumer AI assistants is now measured in months, not years.

OpenAI's GPT-5.6-Cyber found two Chrome zero-days

OpenAI's Daybreak program split into Blue and Red: Daybreak Blue gives approved defenders GPT-5.6 Sol with cyber guardrails removed, while Daybreak Red gates GPT-5.6-Cyber behind stricter vetting for exploit validation and zero-day research. The headline number, a 95% task completion rate versus 1.5% for standard Sol, is a refusal metric: it measures how often the model engages, not how accurate it is.

The concrete win is real. GPT-5.6-Cyber found two previously unknown Chrome V8 vulnerabilities, tracked as CVE-2026-15903, reported them to Google, and both are now patched. Launch partners include Accenture, IBM, CrowdStrike, and Cloudflare. The model is assessed as High under OpenAI's Preparedness Framework, one step below the Critical threshold that paused the Astra model earlier this month.

Security

The "encrypted reasoning" flaw across OpenAI, Anthropic, and Google

Researchers disclosed a flaw in how OpenAI, Anthropic, and Google protect the internal reasoning their models generate: encrypted reasoning objects are portable across sessions, users, and models, so a block created in one session can be replayed into another. In testing, weaker models were used to decode the hidden reasoning of stronger ones. TFTC's analysis reports researchers decoded 315,320 reasoning blocks and recovered live API keys from public agent logs, with 182 credentials extracted from public repositories. All three providers have applied mitigations, but researchers say no architectural fix has shipped, leaving teams that publish session logs exposed. If you run public agents, this is a reason to rotate keys now.

China-linked hackers ran an autonomous AI attack on Taiwan

The Financial Times reported that suspected China-linked hackers used open-source AI agents for what researchers at the Israeli firm Dream call the first fully autonomous, end-to-end cyberattack on a government. Over four days in early July, up to eight AI agents breached a Taiwanese government network, stole data, and compromised accounts, adapting their strategy in real time when blocked and expanding toward nuclear safety and energy targets. SecurityAffairs has the technical breakdown. The attack ran with minimal human oversight, which is exactly the scenario defenders have been warning about since agentic tools went mainstream.

Policy & Industry

DeepSeek builds a team to take on Claude Code

Bloomberg reported that DeepSeek activated official social media accounts around August 12 and posted job listings for a new team building AI agents, specifically a coding tool to rival Anthropic's Claude Code. The company updated its V4 Pro model for autonomous tasks and is recruiting for "agent harness" roles in Beijing. DeepSeek's V4 already runs inside Claude Code; the new unit is an attempt to own the shell, not just the model. The Chinese open-weight lab is moving from model releases to full agent products.

Quick hits

The Takeaway

Three forces converged this week. Open models made agent infrastructure a local commodity. Research agents did real, verified mathematics. And security researchers showed that reasoning traces, the crown jewel of the frontier labs, are not as private as marketed. For builders, the practical read: local agents are now viable, your agent logs are a liability, and the frontier labs are all racing to make agents do the work, not just answer questions. Yesterday's general AI news digest covered the run-up, and the AI news tag tracks the full series.

Related reading from earlier coverage, with the numbers behind each one:

More on AI News

Browse every AI News story →