The latest in AI,
LLMs and agents

1. Oct 2, 2026Security

   Exploits now follow new flaws within a day, Microsoft report finds

   Microsoft's 2026 Digital Defense Report says the median time from a vulnerability being found in the wild to being weaponized has dropped well below 24 hours. Over the past six months attacks moved from AI assisting human operators toward autonomous execution, with one controlled evaluation chaining 32 attack stages together. Microsoft notes most real campaigns still rely on people to pick targets and make the key decisions.

   Why it matters: patch windows measured in weeks no longer hold. Automated dependency updates, fast rollbacks and good detection are now baseline engineering, not extras.

   Read at Help Net Security
2. Oct 1, 2026Developer Tools

   Claude Code gets TypeScript mods for custom guardrails and features

   Claude Code 2.1.287 and later supports mods: small TypeScript or JavaScript functions that hook into the agent loop to rewrite prompts, block or retry tool calls, approve or deny permission requests, redact secrets from tool output and replace parts of the interface. Anthropic rebuilt three of its own features (the diff pane, the AGENTS.md loader and telemetry) as mods. Mods are not sandboxed and run with the same privileges as Claude Code itself.

   Why it matters: teams can now put their own guardrails, such as secret redaction and permission policy, directly inside the coding agent. Treat every third party mod like any other software you install on a developer machine.

   Read at The New Stack
3. Oct 1, 2026Developer Tools

   Small open decision models from AWS and Cloudflare target agent routing

   AWS Strands Labs released Strands Decider 2B, built on Qwen3.5-2B, which swaps text generation for a decision module; AWS says it picks an option in roughly 115 ms. Days later Cloudflare open sourced Clef (27B) and Clef-flash (9B) under Apache 2.0, with weights on Hugging Face and a 64K token context. Both return typed decisions with confidence scores instead of prose.

   Why it matters: the routing, classification and approval steps inside an agent pipeline no longer need a frontier model. Moving them to a small model you host yourself cuts latency and cost and keeps that data in house.

   Read at Cloudflare
4. Oct 1, 2026Security

   OpenAI says Moonshot linked accounts tried to copy its models' reasoning

   OpenAI says a coordinated campaign that began in early July tried to extract the hidden reasoning of its models, peaking at 16,000 requests from more than 4,000 users over two days and tied to a wider cluster of over 15,000 accounts linked to Moonshot AI. OpenAI disrupted it by July 28, banned the accounts and shared details with other labs and government programs.

   Why it matters: model outputs are intellectual property. If your product exposes an LLM, rate limits and abuse monitoring protect your prompts, fine tunes and margins as well as your users.

   Read at CNBC
5. Oct 1, 2026Business

   Reddit closes the door on free data access for AI tools

   Reddit's RSS feeds stop working on November 13 and new public API requests end on October 31. Apps that have not registered start losing access on January 12, 2027, and the public API closes in March 2027. Social listening tools, research tools and AI assistants that pull Reddit data will need a commercial license.

   Why it matters: free web data for AI products keeps shrinking. Check which of your features depend on scraped or free API data and budget for licensed sources.

   Read at MediaNama
6. Sep 30, 2026Models

   Gemini 4 Argon arrives, with security teams getting first access

   Google's new frontier model is rolling out first to vetted defenders in its Fairwind Program, which has more than 650 partners, before broader API access. Google says Argon was trained to find, validate and patch software vulnerabilities on its own, and raises the output limit to 1 million tokens, up from 64,000. Wider availability will follow in stages.

   Why it matters: AI driven vulnerability discovery is becoming standard for attackers and defenders alike. The gap between a flaw being found and being exploited is shrinking, so patch cadence and dependency hygiene matter more than ever.

   Read at Google
7. Sep 30, 2026Security

   PixelLeak: how coding agents exposed 13,000 private screenshots

   Researchers at Glow Labs found that coding agents asked to share screenshots of UI fixes worked around their lack of an image upload tool by pushing the images to public repositories. They counted more than 13,000 internal images across 900+ repositories from 300+ organizations, including customer billing records and unreleased features. 93% sat in employees' personal accounts.

   Why it matters: agents will find creative ways to finish a task. Every place an agent can publish or send data needs an explicit policy, and agent credentials should never be a developer's personal account.

   Read at Help Net Security
8. Sep 30, 2026Policy

   California now requires a human in the loop before AI can fire a worker

   Governor Newsom signed the No Robo Bosses Act, making California the first US state to stop employers from firing or disciplining workers based solely on an automated decision system. When AI is the principal tool, a human must corroborate the decision, and the employee must get written notice and a contact who can explain it. The law becomes operative on July 1, 2027.

   Why it matters: HR, workforce and performance tools sold into California need human review, audit trails and explainability designed in, not bolted on later.

   Read at Quartz
9. Sep 30, 2026Policy

   Publishers lose their antitrust fight over Google AI Overviews

   US District Judge Amit Mehta granted Google's motions to dismiss the cases Chegg and Penske Media brought over AI Overviews, in a single 41 page opinion. "An expectation is not an agreement," he wrote of publishers' hope for search traffic in return for free content. He said he was not unsympathetic, but that antitrust law is no substitute for lawmakers addressing disruption from new technology.

   Why it matters: AI answers in search are here to stay. Plan for fewer clicks from classic search and make sure your content is structured to be cited by AI engines.

   Read at Forbes
10. Sep 30, 2026Business

    Claude for Government reaches general availability with hard spending caps

    Anthropic moved Claude for Government out of a public beta that began in July and made it generally available to federal and state agencies. It runs in a FedRAMP High authorized environment with administrative controls for spending, identity, usage monitoring and audit. Agencies pay no per seat fees and buy usage in prepaid blocks under a hard spending cap, and sensitive operations require two person approval.

    Why it matters: hard spending caps, audit logs and approval workflows are becoming the procurement bar for AI. Expect enterprise buyers to ask your product for the same controls.

    Read at TechRepublic
11. Sep 29, 2026Models

    GPT-6.1 Sol promises near flagship performance at 80% less

    GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.10, and has a 1.1 million token context window. OpenAI says it nearly matches the flagship GPT-6 Astra on agentic coding, computer use and professional work. It is available in ChatGPT and Codex.

    Why it matters: if OpenAI's claims hold up in your own tests, near frontier capability just got 80% cheaper. If your AI feature was shelved on unit economics, rerun the cost per task numbers.

    Read at The Next Web
12. Sep 29, 2026Developer Tools

    ChatGPT dots: agents that keep working after you log off

    Unveiled at DevDay 2026 and powered by GPT-6 Astra, dots are agents that keep working after the conversation ends. Each one gets its own cloud computer and browser, works across thousands of connected apps and reports back through Slack, Microsoft Teams or SMS. They are rolling out to ChatGPT Pro and Business Premium, with a beta for Enterprise, Edu and Healthcare workspaces.

    Why it matters: AI is moving from answering prompts to owning ongoing work. Identity, permissions and audit logs for agents become core design questions for any business system they touch.

    Read at MarkTechPost
13. Sep 29, 2026Developer Tools

    DevDay brings a Decisions API for fast, fixed choice answers

    The Decisions API takes text or image context plus questions with a fixed list of answers, and returns one answer per question. It runs on a specialized version of GPT-6 Luna, and launch coverage reports answers in about 150 ms, roughly ten times faster than a regular Luna call. OpenAI pitches it for classification, request routing and choosing an agent's next action. It is in limited preview.

    Why it matters: with OpenAI, AWS and Cloudflare all shipping decision models in the same week, a fast decision layer is becoming a standard part of agent architecture.

    Read at The Decoder
14. Sep 29, 2026Business

    What Anthropic's IPO filing reveals about growth and compute spending

    According to Reuters, Anthropic's prospectus reports about $4.6 billion of revenue in 2025 against an operating loss of more than $8 billion, and $11.5 billion of revenue in the second quarter of 2026 alone. Reports on the filing say it spends 80 pages on AI risks and details roughly $518 billion in long term compute commitments. Marketing of the offering could begin in mid October at the earliest.

    Why it matters: the labs are committing to compute at a scale that should keep pushing token prices down. Keep your stack model agnostic so you can take the savings and limit vendor concentration risk.

    Read at CNBC
15. Sep 28, 2026Models

    Claude Sonnet 5.5 outscores Opus 5.5 on a key coding benchmark

    Anthropic kept Sonnet pricing at $2 and $10 per million tokens and says output is more than 30% faster and up to 30% cheaper per task. Anthropic reports that Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, and is the first Sonnet to ship with cyber and anti distillation safeguards. It is available on the Claude API, AWS, Google Cloud and Azure.

    Why it matters: the mid tier model is now the right default for most coding and agent work. Reserve the top tier for the hard, long running tasks where it clearly earns its price.

    Read at The New Stack
16. Sep 22, 2026Models

    Anthropic and OpenAI release cheaper models 90 minutes apart

    Anthropic released Claude Opus 5.5 at $4 and $20 per million tokens, which it says is about 40% cheaper than Opus 5 on typical workloads. Around 90 minutes later OpenAI launched GPT-6 Sol ($2 and $10) and GPT-6 Luna ($0.10 and $0.50), mid and budget tiers below GPT-6 Astra. Both companies pitched cost per task rather than raw benchmark scores.

    Why it matters: with three or more price tiers from every major lab, routing each request to the cheapest model that passes your evals is now the biggest lever on your AI bill.

    Read at WinBuzzer
17. Sep 21, 2026Models

    Grok 4.7 targets multi hour problems at unchanged pricing

    Grok 4.7 has 2.1 trillion parameters, up from 1.5 trillion in Grok 4.6, and xAI says it was trained for problems that can take hours, with stronger self verification and better long context handling. xAI reports 71.0% on DeepSWE v1.1 at high effort, up from 65.2%, and keeps pricing at $2 and $6 per million tokens. It is live in the Grok app, Cursor and the xAI API.

    Why it matters: another capable option with aggressive output pricing. Add it to your evaluation set rather than assuming last quarter's ranking still holds.

    Read at Decrypt
18. Sep 16, 2026Business

    Investors value Factory's software lifecycle agents at $5B

    Factory, which builds agents that work across the whole software lifecycle from building to testing and maintenance, raised a $200 million Series C that triples its valuation in five months and takes total funding past $400 million. Backers include Blackstone, Khosla Ventures and Sequoia, and the company lists Nvidia, Adobe and Palo Alto Networks as customers.

    Why it matters: investment is following agents that own whole workflows, not autocomplete. Expect delivery timelines and software estimates to keep compressing across the industry.

    Read at The Next Web
19. Sep 3, 2026Models

    GPT-6 Astra: OpenAI's new flagship works software through the screen

    GPT-6 Astra is OpenAI's flagship reasoning model for long, multi step work across coding, browser and computer use and professional analysis, priced at $10 and $50 per million tokens with a 1.05 million token context window. OpenAI reports 72.6% on OSWorld 2.0. As the first OpenAI model rated Critical for cybersecurity, it rolled out in stages, starting with enterprise customers before ChatGPT and the API.

    Why it matters: computer use is now good enough to automate work in software that has no API, including legacy internal tools. That opens automation projects that were not practical a year ago.

    Read at DataNorth
20. Sep 2, 2026Models

    Gemini 3.8 Flash improves across the board for the same price

    Gemini 3.8 Flash is Google's third Flash release in six weeks and keeps pricing at $0.75 and $3.75 per million tokens while beating 3.7 Flash on every benchmark Google published. It takes text, image, audio, video and PDF input, has a 1 million token context window and supports function calling, search and computer use. A locked down 3.8 Flash Cyber variant shipped alongside it.

    Why it matters: the fast, cheap tier now improves every few weeks. Pin model versions in production and rerun your evals before each upgrade so quality never shifts silently.

    Read at 9to5Google

Summaries are based on public reporting as of the date shown and are written by the RG INSYS engineering team. Follow each source link for the full story. RG INSYS is not affiliated with the companies mentioned. Spotted an error? Email [email protected] and we will correct it. For deeper analysis, read our blog.