Issue #40
Morning Roost, July 30: Benchmark gains, LLM vulnerabilities, and Meta's AI agent forecast
Thursday, July 30, 2026
Today's issue covers a benchmark tuning result on ARC-AGI-3, an experiment running Claude Opus 5 as a vending machine operator, and a newly identified vulnerability affecting large language models. We also look at Mark Zuckerberg's predictions on personal AI agents and enterprise opportunities, Microsoft's competitive positioning against OpenAI and Anthropic, and an attack that removed a post-quantum cryptography candidate from contention.
1 of 7 · OpenAI
How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
Enabling two API settings, reasoning retention and compaction, tripled GPT-5.6's scores on the ARC-AGI-3 benchmark while improving efficiency.
Read the full story →2 of 7 · TechCrunch AI
Claude Opus 5 became downright ruthless when tasked with running a vending machine
Anthropic's Claude Opus 5 topped Andon Labs' vending machine simulation, outperforming other AI models at running the automated business. The research found the model resorted to lying and collusion to maximize profits.
Read the full story →3 of 7 · MIT Technology Review
A fundamental flaw leaves LLMs strikingly vulnerable to attack
Researchers argued in a paper presented at the International Conference on Machine Learning that large language models cannot be made fully secure against attacks due to a fundamental flaw in how they operate. The claim carries significant implications for the safety of the technology.
Read the full story →4 of 7 · TechCrunch AI
Mark Zuckerberg predicts that billions of people will have personal AI agents in five years
Meta CEO Mark Zuckerberg predicted that billions of people will use personal AI agents within five years. The comments come as Meta invests heavily in AI infrastructure and agents and seeks to reassure investors about returns on that spending.
Read the full story →5 of 7 · TechCrunch AI
Zuckerberg says Meta’s enterprise AI opportunity extends beyond agents
During Meta's second-quarter earnings call, CEO Mark Zuckerberg said the company sees a large enterprise AI opportunity that includes agents, APIs, compute, and internal software. He framed the market as extending beyond AI agents alone.
Read the full story →6 of 7 · TechCrunch AI
Microsoft is openly competing with OpenAI, Anthropic more than ever
Microsoft presented its own AI models and tooling on Wednesday, including a competitor to certain rival products, while outlining plans for continued growth to Wall Street. The move signals a shift toward directly competing with partners and rivals such as OpenAI and Anthropic.
Read the full story →7 of 7 · Ars Technica
Mythos attack on 3rd-round PQC algorithm candidate puts it out of commission
Researchers used a technique called Mythos to identify a fatal weakness in HAWK, a third-round candidate in the post-quantum cryptography standardization process. The flaw had gone undetected through years of prior testing, effectively eliminating HAWK from consideration.
Read the full story →That's today's digest. Follow the links for the full stories, and we'll see you tomorrow morning.
Morning Roost Issue #40