U.S. CPI: +2.9% YEAR-OVER-YEAR (EST. +2.9%); U.S. CORE CPI: +3.2% YEAR-OVER-YEAR (EST. +3.3%)

Beating AI Express: Google said its Gemini AI model accidentally accessed the internet and gained access to the systems of three companies during a cybersecurity capability test, marking the first known case of a Google AI system autonomously performing such an operation. The incident occurred in May during a test conducted by cybersecurity firm Irregular. During the test, Gemini was performing tasks in a "capture the flag" (CTF) cybersecurity exercise, but it accessed real corporate systems because the virtual company names in the test environment matched real ones, plus the model was accidentally granted internet access. Google stated that in one instance, the model accessed a protected system by attempting passwords; in the other two cases, it found credentials in public online repositories and tried to access the corresponding systems. After confirming it had accessed real corporate systems, the model halted all further actions. Google noted the incident caused no harm to the affected companies, so it was not classified as an AI model out-of-control incident, adding that safety mechanisms helped the model stop its actions in time. The affected companies have been notified, and Google has also reported the incident to relevant regulatory authorities. The incident is similar to previous security test events disclosed by AI firms including OpenAI and Anthropic, and has once again drawn external attention to AI models’ ability to autonomously perform network operations and their safety protection mechanisms.

Beating AI News Flash: According to Reuters, three people familiar with the matter revealed that ahead of Anthropic’s upcoming IPO, and following the company’s CEO’s call for the entire industry to slow its pace, Anthropic is considering launching a new AI model to counter the momentum OpenAI has built since releasing GPT-6Astra. OpenAI’s GPT-6Astra, launched this month, has garnered widespread attention among enterprises, sparking concerns among some investors. They told Reuters that they are re-evaluating Anthropic’s standing as a leading provider of enterprise AI tools. A source familiar with the matter said Anthropic is assessing the safety of its next-generation model as one of the key considerations for whether to release it. Another source with insight into the company’s thinking noted that some discussions center on balancing investments in the new model’s launch and boosting profitability, as rising interest rates have made investors more focused on the timeline for achieving expected profits. Separately, two people familiar with the matter disclosed that Anthropic may delay its IPO until after the U.S. midterm elections in November, with expectations that the election will not significantly impact the offering. The IPO has already been pushed back from its original schedule; Reuters previously reported that roadshows were anticipated to start as early as mid-October.

Beating AI News Insight: Google confirmed that its Gemini model accessed the systems of three real companies during a cybersecurity evaluation. One entry was gained via password guessing, while the other two obtained login credentials from public code repositories. Gemini halted operations immediately after determining the targets were real companies. No sandbox escape occurred throughout the process, as the public network entry was already open by default. The root cause lies with third-party evaluation firm Irregular: the test was not supposed to access the public network, but an environment misconfiguration allowed the model to connect directly to the internet. The simulated companies used names identical to real ones, leading Gemini to trace the task to external real-world systems. Irregular stated that this is the same issue underlying similar incidents at other labs in the past. Anthropic, OpenAI, and Meta all encountered similar pitfalls during Irregular’s evaluations previously. Both Anthropic and OpenAI confirmed the incidents stemmed from the public network being mistakenly enabled in the test environment; OpenAI further emphasized that no complex sandbox escape or zero-day exploit was involved. Independent media outlet Effort therefore criticized that these incidents were later framed as "model runaway" narratives.

Beating AI News Flash — According to a civil lawsuit filed Friday in federal court, AI giants Anthropic, OpenAI, SpaceX AI, and Google face conspiracy charges over their recent calls to coordinate a slowdown in AI development. The lawsuit alleges that Anthropic CEO Dario Amodei publicly called for "industry-wide coordination" earlier this month to "set the pace for cutting-edge AI" — a move endorsed by SpaceX AI head Elon Musk, OpenAI CEO Sam Altman, and Google DeepMind co-founder Demis Hassabis — which amounts to an illegal business agreement between competitors under U.S. antitrust law. Nick Rowley, one of the attorneys representing the four plaintiffs in the U.S. District Court for the Northern District of California, stated that the case aims to ensure secret, self-serving agreements among the world’s most powerful for-profit tech companies do not lead to AI "quickly spiraling out of human control." When it comes to threats of existential events like nuclear war and the largest risk in human history today, people deserve ironclad safeguards. The rule of law should be established transparently, legally, and accountable to the public by the U.S. government.

Beating AI Express News: Anthropic has added AGENTS.md support to Claude Code. Starting from version 2.1.277, Claude Code will automatically read AGENTS.md if a project lacks a CLAUDE.md file; users can also configure the tool to load both files via the /config setting. AGENTS.md is a cross-tool standard project rule file already supported by tools including Codex, Cursor, Gemini CLI, and GitHub Copilot. This is exactly the issue Shopify CEO Tobi Lütke publicly flagged at the end of August. Shopify uses multiple AI agents including Claude Code and Codex; prior to this update, Claude Code only recognized CLAUDE.md, forcing the team to maintain two separate sets of rules. In large codebases, if synchronization is missed at any layer, different tools may end up with conflicting rules. Tobi even stated at the time that he was considering disabling Claude Code at Shopify. Back then, Thariq Shihipar, a Claude Code engineer, explained that Anthropic had stuck with its own CLAUDE.md due to different models being suited to different prompts, skills, and project instruction formats. Less than a month later, Anthropic has rolled out AGENTS.md support as a built-in mod for Claude Code, and made its source code public. Going forward, developers can also customize how project instructions are loaded based on this mod.

The American Association of Individual Investors (AAII) weekly survey shows investor sentiment has worsened further to its lowest level since 2025, fueled by sustained high oil prices and the Federal Reserve’s first interest rate hike since 2023. The latest AAII survey, released Friday, found 53.3% of respondents are bearish, exceeding the 50% mark. Just 28.8% are bullish, the lowest such reading in a year. In a statement, AAII Vice President Charles Rotblut said the gap between bullish and bearish investors has dropped to -24.5%, a level he described as “abnormally low” and marking the ninth straight week it has stayed below the historical average of 6.5%. This is also the metric’s lowest level since May 2025. The survey further shows more than half of respondents hold higher-than-normal cash allocations, indicating investor caution, with 19.1% reporting their cash holdings are “far above normal levels.”