Noticiário do Joãozinho: Edition no. 17 · Nov. 8, 2026 · Morning

Medium importanceNov. 6, 2026

OpenAI brings in Hacktron to hunt for flaws in its own systems, with expanded access to Daybreak Red

OpenAI announced (11/06) a partnership with Hacktron, which builds AI agents for code review, penetration testing and red teaming, to find vulnerabilities in OpenAI's systems, focusing on how they interact with third-party services. Engineers from both companies will jointly choose the areas to test, and Hacktron gets more access to frontier models through Daybreak Red. OpenAI met the company through its bug bounty program, whose rewards were recently increased.

Medium importanceNov. 5, 2026

Palo Alto Networks in talks to buy Clutch, which protects AI agents' credentials, for US$ 500 million

Palo Alto Networks is in talks to buy Israeli startup Clutch Security for US$ 500 million in cash, plus milestone payments, according to a TechCrunch source (The Information reported it first). Three-year-old Clutch monitors and protects the passwords, tokens and keys that AI agents use to access company systems. It was valued at US$ 85 million in its January 2025 Series A and is approaching US$ 10 million in annual recurring revenue, which would mean about 50 times ARR. Neither company commented; the deal may not happen.

Medium importanceNov. 8, 2026

AP: staff at the federal AI safety center now report to the White House

According to AP (11/08), staff at CAISSI (Center for Advancing Innovation and Standards for Super Intelligence), the Commerce Department division that assesses the safety of new AI systems, were told to report informally to a White House office, according to two former government officials. The center has had no permanent director since the last one left, has stopped publishing its test reports and has a budget of US$ 10 million a year, about one-eighth of its UK counterpart. The White House did not answer questions about CAISSI.

Medium importanceNov. 5, 2026

NVIDIA opens four Cosmos 3 Nano models to simulate driving and robots in real time

NVIDIA published on Hugging Face four open-weight checkpoints of the Cosmos 3 family, all derived from Cosmos 3 Nano with 16 billion trainable parameters and the OpenMDW1.1 license (commercial use allowed): Transfer-Auto (driving data with up to 11 cameras and LiDAR, varying weather and lighting), Sim-Auto (distilled version that generates camera and LiDAR in streaming), Sim-Transfer (turns a physics simulator's controls into realistic video) and Sim-Bimanual (predicts video from the actions of a two-armed robot). They come with post-training recipes and “skills” for coding agents.

Medium importanceNov. 2, 2026

GitHub Actions will block pull_request_target by default on 12/02, with the first test this Monday (11/09)

Starting 12/02/2026, GitHub Actions blocks pull_request_target events in all public repositories unless an event policy allows them. The date was 11/02 and was postponed. Before that there are two “brownouts”, with the block actually switched on: 11/09 and 11/16, from 16:00 to 18:00 UTC (1 p.m. to 3 p.m. Brasília time). The reason: this trigger runs in the context of the base repository, with access to secrets, and untrusted code from a fork can steal them. Private and internal repositories and GitHub Enterprise Server are not affected.

Medium importanceNov. 6, 2026

Nous Research releases Arete-27B: Qwen3.8-27B retrained to remove Chinese government bias, with open weights

Nous did supervised fine-tuning of Alibaba's Qwen3.8-27B with about 52.8 million tokens, using Inkling as the teacher. On its own evaluation of 803 questions across seven topics, answers with propaganda fell from 22% to about 1% and the fact score rose from 2.32 to 2.98 (out of 5). On Terminal-Bench-2.1 it scored 70% versus 67% for the base model, as run by Nous. The weights are on Hugging Face under Apache 2.0 (54.7 GB in BF16), and there are already community GGUFs: Q4_K_M is 17.2 GB.

Medium importanceNov. 4, 2026

OCC gives examiners a guide to assess AI and agents at US banks

Bulletin 2026-54 from the OCC, the regulator of US national banks, brings a supervisory framework for the use of “super intelligence” (the term the government now uses for AI). It creates no new requirements: it tells examiners to tailor supervision to the materiality of each use. Key points: AI is not a separate risk category; the intensity of controls should grow with the consequences of an agent's failure or misuse; and risk classification should consider consequences, how much autonomy was delegated and the degree of human oversight, not the model's technical sophistication.

Medium importanceNov. 7, 2026

JPMorgan's CISO suggests taking services off the open internet because of attackers' AI agents

Pat Opet, chief information security officer at JPMorgan Chase, wrote on X (11/07) that people working in cybersecurity should be evaluating how to “abstract your companies services from the internet”: “why allow an unauthenticated anonymous agent of an attacker to interact with your systems?”. Greg Brockman, OpenAI's president, quoted the post saying security architecture fundamentals “are more important than ever before”, and investor Ed Sim summed up the idea as “private by default”. It is architecture advice, not a product announcement or a rule.

Medium importanceNov. 6, 2026

SEGA Sammy reaffirms: no generative AI directly in the creative side of games

In its Integrated Report 2026, external director Kohei Katsukawa says he “strongly” supports CEO Haruki Satomi's policy of not using generative AI directly in the creative aspects of games and other content, because valuing human creativity builds trust with fans. The same report says nearly all group employees in Japan already use generative AI day to day and that it remains allowed for development support, operations and marketing, under copyright and data guidelines. The passage was highlighted by the Sonic City site on 11/06.

Medium importanceNov. 6, 2026

Brazilian government allocates R$ 46.3 million to CPQD for a national data center network that distributes tasks with AI

Finep signed on 11/06 the release of R$ 46.3 million from Funttel for the Prisma project, by CPQD (Campinas), with a 36-month timeline. The platform will connect data centers across the country, including small ones in the interior, and use AI to send each processing task to the center with the best energy efficiency and capacity at that moment, and can even move operations already under way. The proposal cites digital sovereignty, lower energy use and more resilience to outages and attacks.

XMedium importanceNov. 8, 2026

Addy Osmani: “your coding agents made you a manager”

In an article on X, Addy Osmani, who works on Claude Code, compares good engineering manager practices with using agents: give the agent a way to check its own work (tests, build, screenshot), explain what it cannot see in the repository, record in CLAUDE.md anything corrected twice, restart the session after two failed corrections and review every diff “like you're going to get paged for it”. He also points out where the analogy breaks, for example: running five agents does not prepare anyone to manage five people.

Engagement 71.9K views·1.2K likes·124 reposts·48 quotes·108 replies·1.9K bookmarks (as of Nov. 8, 2026 08:17)