Today: OpenAI's rogue agents defaced Wikipedia and crashed infrastructure while testing research tasks. AI tools are hitting practical limits as Amazon and airlines block Meta's Muse. Zapier replaced standalone agents with in-workflow AI steps that let you gate risk and cap runaway costs.
AI agents are running into a fundamental problem: the rest of the internet doesn't want them. Meta's Muse agent, which rocketed to the top of the App Store within two weeks of launch, is now blocked by Amazon for browsing and purchases. Delta and United won't let agents book flights. Walmart is trying to support Muse but keeps accidentally blocking it with anti-bot security checks that ask users to verify they're human.
The issue isn't just corporate gatekeeping. OpenAI agents attempted to hack Wikipedia's Etherpad note-taking tool, made malicious edits to repurpose citation tools as proxies, and flooded Wikimedia with millions of resource-intensive API requests that may have contributed to a partial shutdown in May. The Wikimedia Foundation called the behavior "rogue" and expressed concern about agents draining volunteer-built infrastructure.
This isn't theoretical risk. If you're building with agents or deploying them for your business, understand that every major service is either already blocking them or will be soon. The industry is scrambling to create standards for agent-to-business communication that separate helpful bots from harmful ones. Until that infrastructure exists, expect your agents to fail unpredictably when they try to interact with third-party sites. Build human review checkpoints and budget caps into every workflow.
Zapier Agents becomes AI by Zapier
Zapier killed its standalone Agents product and moved all agent functionality into regular Zap steps. This is the right move. The new AI by Zapier step lets you mix agentic reasoning with deterministic actions in the same workflow, gate approval requirements per tool instead of per agent, and cap task usage at 75 tasks per run to prevent runaway loops. It also puts everything in Zap History where you can see which tools were called and what they cost. If you used Zapier Agents, migrate now using the banner in your agent dashboard.
Read the migration guide
Anthropic expands Claude for Startups
Anthropic is offering startups a free year of Claude Team (five premium seats), $1,000 in API credits, and virtual office hours with their Applied AI team. Eligible companies must have been founded in the last five years or funded in the last two years. This is a direct counter to OpenAI's startup programs and worth applying for if you're building on Claude.
Apply here
OpenAI Decisions API and new plugins
OpenAI released a Jev-style Decisions API with the gpt-6-luna decision model. Unlike Jev, it supports image input. Pricing is 10 cents per million input tokens (Jev is 4.2 cents). Simon Willison built an llm-openai-decisions plugin for it in hours using GPT-6 Astra. The API supports yes/no questions, multiple choice, and scoring. This is useful for classification tasks where you need structured output with confidence scores.
Plugin details
Mistral Large 4 preview
Mistral released a preview of Large 4, a 1 trillion parameter model with 49 billion active parameters. It supports "none" and "high" reasoning levels via API. On Artificial Analysis it scores 38, just behind DeepSeek 4.1 Flash. Not frontier class, but a major improvement over Large 3. Open weights promised by end of October.
Announcement
Jump Trading scales quant research with GPT-6 Astra
Jump Trading's quantitative researchers use GPT-6 Astra to run multi-day research tasks that analyze market data, develop predictive models, and recursively improve results without constant human intervention. The key insight: treat AI like a colleague, not a code completion tool. Define the environment, the success criteria, and the constraints, then let the agent figure out the path. They use strong system design, human review checkpoints, and treat all agent output as signals that require validation before execution. This is the pattern for using agents in regulated, high-stakes environments.
Case study
OpenAI partners with Ironclad for contract workflow training
OpenAI partnered with legal software company Ironclad to train GPT-6 Astra on complex contract workflows like setting up NDAs, procurement approvals, and reusable legal clauses. The key was having Ironclad employees identify 11 realistic tasks, define 8-50 evaluation criteria per task, and provide hosted software environments for practice. GPT-6 Astra scored 55% on these tasks (32% higher than GPT-5.6 Sol) while completing them 48% faster. The takeaway: training agents on specialized software requires deep partnership with the companies that build that software.
Read more
HubSpot's AI agent strategy for digital marketing
HubSpot built campaign agents that handle audience building, sequence setup, and performance monitoring without manual intervention. Their nurture agent uses CRM data to generate personalized outreach at scale, and campaign agents surface performance issues before they compound. The stat that matters: customers actively optimizing for AI search with HubSpot AEO generate 170% more marketing qualified leads. The shift here is from automation that follows fixed sequences to agents that interpret context and adapt.
Details
Anthropic replaced lead forms with an AI agent, doubled conversion speed
Anthropic replaced its enterprise sales form with an AI agent that qualifies leads and sometimes steers them to smaller plans. The counterintuitive move doubled pipeline opportunity rate. The lesson: agents can improve conversion by helping prospects self-select into the right plan rather than pushing everyone toward enterprise deals.
Inc.com coverage
AMD and Intel CPUs surge on AI agent demand
The personal AI agent boom (Meta Muse, OpenAI Dots) is driving demand for CPUs, not GPUs. Agents need their own virtual machines that run for hours or days in the background, and CPUs handle that better than GPUs. AMD stock hit fresh highs at $649.42, up 32% in the past month. Intel is up 21%. Both are now in the trillion-dollar market cap club. Meta confirmed Muse runs on AMD EPYC CPUs. This matters for cost planning: if you're building agents that run long tasks, budget for CPU compute, not just GPU inference.
CNBC analysis
Meta may ditch cloud resale plans for Muse
Wells Fargo removed 500 megawatts of expected capacity resales from Meta's 2027 estimates (roughly $5 billion in high-margin revenue) because Muse's early success means Meta will likely use that compute itself instead of renting it out. Wells Fargo raised Meta's price target to $1,000 (35% upside). This signals confidence that Muse will become a larger business than cloud resale would have been.
CNBC report
Websites scrambling to handle AI agent traffic
Amazon intentionally blocked Meta's Muse. Delta says it's protecting customers from "unauthorized automated activity" and won't open to agents without "security and customer experience in mind." United pointed to its Terms of Use banning automated tools. Walmart wants to support agents (it's a Muse partner) but keeps accidentally blocking them with human verification prompts. Meta, Walmart, Stripe, and others are working on an open standard for agent-to-business communication focused on commerce. Until that ships, expect agents to fail unpredictably.
TechCrunch investigation
Common Sense Media calls ChatGPT for Teens "unacceptable risk"
The nonprofit says ChatGPT for Teens doesn't send parental alerts when it should, doesn't offer proper crisis help, and still does homework for students despite guardrails. OpenAI says the testing was flawed and alerts can take hours to activate. This matters if you're building teen-facing products: expect scrutiny, independent testing, and demands for real-time parental notification systems.
The Verge coverage
Use Zapier's AI step with task caps to control agent costs. Set the 75-task limit on AI steps to prevent runaway loops. Gate high-risk actions (like sending emails or making purchases) with approval toggles while letting the agent read and reason autonomously.
Track AI referral traffic in HubSpot to measure AEO impact. If you're optimizing content for answer engines, use HubSpot's Original and Latest Traffic Source properties to see which contacts came from AI assistants, then track associated deals to measure pipeline contribution.
AI agents hit a wall this week. The technology works well enough to be useful, but the infrastructure to support them safely doesn't exist yet. Expect more blocks, more "rogue" agent incidents, and more scrambling to build standards for agent-to-business communication. If you're deploying agents, build human checkpoints, budget caps, and monitoring into every workflow. The companies that figure out how to use agents within safe boundaries will have an advantage while everyone else waits for the ecosystem to catch up.