DEV Community: Felipe L

ChatGPT now tracks your browsing via ad collector

Felipe L — Sun, 20 Sep 2026 22:19:32 +0000

What Happened

ChatGPT now includes an ad‑collector that reads the content of other websites you visit. The update, announced in a recent blog post, says the feature is meant to sharpen ad targeting and personalization. Practically, as you browse, ChatGPT’s system can access the pages you view, capture text, and use that data to tailor responses or ads.

The ad‑collector builds on existing web‑retrieval capabilities, but it is specifically tuned to read browsing history for advertising purposes. This raises immediate concerns about user privacy, data ownership, and how the collected information is stored or shared.

Why This Matters for Builders

Data‑handling policies: If your automation or AI‑agent workflow pulls user data from external sites, the ad‑collector may intercept that data. This could conflict with GDPR, CCPA, or internal policies.
Compliance and audit trails: Regulated environments need clear audit logs of data origin and processing. The ad‑collector introduces a source that may not be easily auditable, creating compliance gaps.
Trust and user experience: Users expect the AI you build to use only data you explicitly provide. Knowing that ChatGPT can read their browsing history can erode trust, especially if your workflow collects or analyzes that data.
Workflow design: If your n8n or other platform integrates with ChatGPT, add safeguards—such as opt‑in prompts or data‑masking steps—to avoid feeding the model private browsing content.
Security posture: The ad‑collector could leak data. If your workflow passes user data to ChatGPT, verify that no unintended information is transmitted.
Competitive differentiation: Some organizations may avoid using ChatGPT for certain use cases to maintain higher privacy. Knowing this feature helps you decide whether to switch to an open‑source model or a provider with stricter controls.

FAQ

Q: Can I opt‑out of the ad‑collector when using ChatGPT in my workflow?

A: The current implementation does not allow a user‑controlled opt‑out. You must manage data flow at the application level and send only the data you explicitly intend to share.

Q: Will the ad‑collector affect data already stored in my database?

A: The ad‑collector reads real‑time browsing data. It does not scan stored data unless you feed that data into ChatGPT. Any data you send could combine with browsing insights, so review your sharing practices.

Q: How should I update my compliance documentation to account for this feature?

A: Add a section that explains the potential for third‑party data capture via ChatGPT’s ad‑collector, outline mitigation steps (e.g., data minimization, explicit user consent), and document how you monitor and audit data flows involving the model.

Originally published on Automations Cookbook.

AI‑slowdown collusion lawsuit hits OpenAI, Anthropic

Felipe L — Sun, 20 Sep 2026 18:03:57 +0000

What Happened

OpenAI, Anthropic, and other AI firms face a federal antitrust lawsuit. The U.S. Department of Justice alleges they formed an illegal agreement to slow the rollout of advanced AI models. The suit claims the companies coordinated release schedules and shared model‑capability data to avoid competition and regulatory scrutiny. Early evidence points to collusion that could hurt consumers and developers who depend on rapid AI improvements.

Why This Matters for Builders

Regulatory uncertainty: New oversight could force automation teams to audit AI‑agent pipelines. Expect requirements for more transparency, detailed logging, and audit trails.
Model availability and stability: A slowdown in releases would mean fewer updates. Builders should plan for longer upgrade cycles and create fallback strategies.
Cost implications: Penalties or altered pricing models could raise API costs. Evaluate cost‑benefit trade‑offs and consider alternative or hybrid providers.
Innovation pacing: If progress was intentionally throttled, future breakthroughs may be delayed. Teams relying on cutting‑edge features should adjust roadmaps and invest in internal research or open‑source options.

FAQ

Q: Will this lawsuit stop the release of new AI models?

A: Not necessarily. Firms may continue releases while the case proceeds, but they might adopt more cautious strategies to reduce legal risk.

Q: How should I prepare my n8n workflows for potential regulatory changes?

A: Add comprehensive logging, version control, and audit trails to all AI calls. Keep detailed records of model versions, usage metrics, and compliance checks.

Q: Could I switch to open‑source models to avoid these risks?

A: Open‑source models reduce vendor‑specific regulatory exposure, but they require maintenance. Weigh the trade‑off against your team’s capacity and risk tolerance.

Originally published on Automations Cookbook.

Microsoft Director Calls AI Scraping the Largest Theft of Labor in Hu…

Felipe L — Sun, 20 Sep 2026 12:44:06 +0000

What Happened

Microsoft’s director of research and AI called AI scraping—the practice of feeding large language models millions of copyrighted texts without permission—the biggest theft of labor ever. The comment came as lawsuits target OpenAI’s ChatGPT and other generative models for copying copyrighted content without paying creators. The director pointed out that AI models now train on billions of web documents, many protected by copyright. The industry says this is “fair use” or essential for useful models, but the warning suggests legal and ethical rules may lag behind technology. The debate shifts to whether developers should license content, give attribution, or find new ways to honor intellectual property.

Why This Matters for Builders

Data provenance and compliance: If your automation or AI‑agent workflow pulls web or internal data, verify that the source is licensed for reuse. Unchecked scraping can expose your organization to legal risk, especially when that data feeds downstream AI services.
Model training and fine‑tuning: When building custom models or fine‑tuning existing ones, check whether the training data contains copyrighted material. Curate datasets, add copyright notices, or use public‑domain corpora to avoid infringement.
Ethical workflow design: Builders must design agents that respect creators’ rights. Add content filters, include attribution in outputs, or flag potential copyright violations.
Operational resilience: Legal challenges can disrupt data and model APIs. If a provider faces litigation, your automation pipelines may suffer. Build redundancy and monitor policy changes to mitigate risk.

FAQ

Q: Should I stop using web‑scraped data in my AI agents?

A: Not automatically. Audit the data for licensing and consider licensed or public‑domain sources. Add a compliance layer that checks copyright status before ingestion.

Q: How can I add attribution to outputs from my AI agents?

A: Capture source metadata during collection and embed it in the agent’s response. Append a citation string or generate a structured attribution object.

Q: What if my workflow uses third‑party APIs that might be scraping data?

A: Review the API’s terms of service and data usage policies. If the provider is known to scrape copyrighted content, switch to a licensed‑data service or build your own pipeline with clear provenance.

Originally published on Automations Cookbook.

CUA‑S1: System One Model for Computer Use

Felipe L — Sat, 19 Sep 2026 22:16:31 +0000

What Happened

The CUA‑S1 project on GitHub introduces a System One model that unifies multiple tools into one workflow, driven by a lightweight AI agent. Users can issue high‑level commands, and the system translates them into actions across the underlying stack. The proof‑of‑concept implementation connects to productivity tools, file systems, and cloud services using a minimal set of APIs. It shows the AI agent orchestrating tasks such as data retrieval, processing, and dashboard updates automatically.

Why This Matters for Builders

Fewer context switches: Automation teams can keep the agent focused on a single mental context, which boosts reliability and simplifies debugging.
Single API surface: Exposing one endpoint for several downstream services cuts wrappers and adapters, easing integration.
Lower operator load: Operators no longer need to remember which tool performs which task; the AI abstracts that, speeding onboarding and cutting errors.
Scalable workflow composition: Adding new steps or services becomes a matter of extending the model, not re‑engineering pipelines.
Consistent security posture: Centralizing permissions and audit logs reduces compliance complexity and shrinks the attack surface.

FAQ

Q: Can I integrate CUA‑S1 with my existing n8n workflows?

A: Yes. The model offers a REST API that n8n can call, letting you trigger System One actions within your flows.

Q: What kind of AI model powers System One?

A: The project is LLM‑agnostic. It works with open‑source models like Llama or proprietary APIs such as OpenAI, as long as they understand the command schema.

Q: How does System One handle failures or partial results?

A: Built‑in retry logic and a state machine track progress. This allows graceful recovery and clear error reporting to downstream systems.

Originally published on Automations Cookbook.

OpenAI Uses LLMs to Design Its Own Jalapeño Chip

Felipe L — Sat, 19 Sep 2026 17:49:55 +0000

What Happened

OpenAI used its own large language models to design a new custom chip called jalapeño. The team fed architectural constraints and performance goals into an LLM trained on code and design data. The model generated RTL (register‑transfer level) descriptions, simulated them, and iterated until the design met specifications. The chip, optimized for OpenAI’s workloads, was fabricated and integrated into the infrastructure.

This proof‑of‑concept showed that a generative AI model can replace, or at least cut, the manual effort needed for ASIC and FPGA design. Automating synthesis and validation shortened the design cycle from months to weeks, enabling rapid prototyping of accelerators for new AI workloads.

Why This Matters for Builders

Accelerated Prototyping: LLM‑driven design shrinks hardware development time, letting builders iterate on custom accelerators for niche workloads without waiting for a full silicon cycle.
Reduced Expertise Barrier: Teams lacking deep hardware experience can still experiment with custom chips. LLMs translate high‑level specs into RTL, lowering the skill threshold.
Integration with Automation Workflows: The same LLMs can be added to existing n8n or AI‑agent pipelines to generate configuration files, deployment scripts, or low‑level code snippets, creating a unified AI‑driven engineering stack.
Cost Efficiency: Faster cycles mean fewer engineering hours and silicon iterations, saving money for startups and mid‑sized companies building AI infrastructure.
Future‑Proofing: As AI models evolve, LLM‑based design allows continuous re‑optimization of chips to match new architectures, keeping infrastructure ahead of demand.

FAQ

Q: Can I use an LLM to design my own custom AI accelerator?

A: Yes. With a powerful language model and a dataset of hardware design examples, you can generate RTL or high‑level synthesis code. Hardware verification expertise remains necessary to validate the output.

Q: Will LLM‑generated hardware be as reliable as traditional designs?

A: Initial outputs often need human review and refinement. Reliability will grow as models improve and more design data is used, but verification is still essential.

Q: How does this fit into my existing n8n automation workflows?

A: You can build a workflow that sends design constraints to an LLM endpoint, receives RTL, runs it through a synthesis tool, and triggers deployment scripts—all orchestrated by n8n nodes for full automation.

Originally published on Automations Cookbook.

Anthropic Adds AGENTS.md Support to Claude Code

Felipe L — Sat, 19 Sep 2026 12:08:31 +0000

What Happened

Claude Code now supports AGENTS.md, a markdown format for defining AI agents. Developers can list an agent’s capabilities, prompt templates, and workflow logic in one human‑readable file. The change lets Claude read the file at runtime, so agent behavior updates without redeploying code. Linting and validation tooling ensures the markdown follows a schema Claude Code can parse. Anthropic announced the update in a brief tweet, but the impact on automation builders is significant.

Why This Matters for Builders

Declarative agent design – Writing agent logic in markdown keeps behavior in version control with other infrastructure code, boosting traceability and auditability.
Rapid iteration – A single file change can instantly redeploy a prompt or workflow, shortening the A/B testing and fine‑tuning cycle.
Cross‑platform compatibility – AGENTS.md is language‑agnostic. Whether you use n8n nodes, Zapier actions, or custom scripts, the same markdown feeds Claude Code, reducing onboarding friction.
Reduced boilerplate – The format abstracts repetitive prompt engineering and state management, letting builders focus on business logic.
Better collaboration – Plain‑text files let non‑technical stakeholders review and suggest changes, easing product‑manager or data‑scientist involvement in agent design.

Teams already using Claude can migrate existing agents to AGENTS.md without rewriting their codebase. The update also enables adding Claude agents to n8n workflows via a node that reads AGENTS.md and passes it to Claude Code.

FAQ

Q: Does AGENTS.md require a new version of Claude Code?

A: Yes. The feature is available only in the latest Claude Code release; update your environment to use it.

Q: Can I use AGENTS.md with existing n8n nodes?

A: Create a custom node that reads the markdown file and forwards its content to Claude Code, integrating seamlessly with current workflows.

Q: Is there a limit to the size of AGENTS.md?

A: The file is parsed in memory. Very large files may affect performance; keep the file focused on core agent logic.

Originally published on Automations Cookbook.

Claude Code now reads AGENTS.md if Claude.md missing

Felipe L — Fri, 18 Sep 2026 22:26:03 +0000

What Happened

Claude Code now falls back to an AGENTS.md file when a project lacks Claude.md. Agents defined in AGENTS.md load automatically, eliminating the need for a placeholder file. The changelog noted this update as a fix for configuration errors.

Why This Matters for Builders

Less manual setup: Skip creating a dummy Claude.md when all agent logic lives in AGENTS.md. Saves time in CI/CD pipelines.
Fewer runtime failures: Missing Claude.md used to stop agents from loading. The fallback keeps them discoverable and prevents silent production crashes.
Cleaner version control: A single source of truth (AGENTS.md) reduces merge conflicts and simplifies history in large teams.
Predictable tooling: Automation platforms like n8n can rely on one file name, streamlining scripts and deployment pipelines.

FAQ

Q: Does this change affect existing projects that already have Claude.md?

A: No. If Claude.md exists, the system uses it. The fallback only triggers when that file is absent.

Q: Will the fallback work with nested agent definitions or only top‑level ones?

A: The system parses the entire AGENTS.md file, so any agent definitions—nested or flat—are loaded.

Q: How should I update my deployment scripts to take advantage of this?

A: Reference AGENTS.md instead of Claude.md in your build or deployment steps. If you still create a Claude.md, the system will prioritize it, so pick one convention for consistency.

Originally published on Automations Cookbook.

GrassLobster: AI Agentic Generation of Parametric Geometry Workflows

Felipe L — Fri, 18 Sep 2026 18:07:34 +0000

What Happened

GrassLobster appears in a new Miro Vision feature. The agent generates parametric geometry workflows from a short description, like “create a 3‑story modular building with variable floor plans.” It outputs a fully‑functional workflow that runs in a design environment. The system uses large language models for natural‑language understanding and a geometry‑aware inference engine to convert intent into parametric constraints, nodes, and dependencies.

A public demo lets users adjust room count, window placement, or structural load. The agent instantly regenerates the workflow. Designers can export the output to popular tools and refine it further. The demo shows AI handling rule‑based geometry while humans make creative choices.

Why This Matters for Builders

Rapid Prototyping: Integration with automation platforms like n8n can spin up design workflows in hours instead of days.
Consistent Quality: Codified design rules reduce human error and enforce regulatory and performance standards.
Scalable Customization: A simple UI can feed user preferences into the agent, letting non‑technical stakeholders influence design without coding.
Seamless Integration: The agent’s output can be wrapped as a node in an AI‑agent workflow, creating end‑to‑end pipelines from requirement capture to CAD export, all orchestrated by an automation engine.
Cost Efficiency: Automating bulk geometry logic frees designers to focus on value‑added tasks, boosting productivity and cutting labor costs.

FAQ

Q: Can GrassLobster’s output be used directly in n8n workflows?

A: Yes. The agent exposes a REST endpoint that returns workflow definitions in JSON. n8n can fetch this JSON, parse it, and instantiate the corresponding nodes, making integration straightforward.

Q: How reliable is the geometry generated by the agent?

A: The agent follows design rules encoded in its training data, but human review still helps. A typical workflow includes a validation step that checks constraints before final export.

Q: What kind of performance impact should I expect when adding this to a production pipeline?

A: Inference time is comparable to other LLM‑based services—generally a few seconds per request. For large batches, consider asynchronous execution or caching results to keep latency low.

Originally published on Automations Cookbook.

Microsoft, OpenAI Lose Fight to Hide Internal Docs on AI Scraping

Felipe L — Fri, 18 Sep 2026 12:27:04 +0000

What Happened

Microsoft and OpenAI were forced by a lawsuit from tech‑policy advocates to hand over internal documents. Those documents reveal that the companies scraped massive amounts of web text. The court ruled that the scraping constituted a theft of labor and called it the largest theft in human history. The firms must stop the practice and submit a detailed audit of the data used in their models.

The ruling puts any AI model built with similar scraping techniques under legal scrutiny. Regulators are now watching the industry more closely, and the precedent could reach other large AI vendors.

Why This Matters for Builders

Data sourcing compliance: If your n8n or AI‑agent workflow pulls data from public sites, verify that the source is licensed or that you have explicit permission. Relying on scraped data exposes you to liability.
Model retraining and data freshness: The ruling may force vendors to abandon scraped corpora. You’ll need fresh, properly licensed data or open‑source models trained on public datasets.
Auditability and transparency: Automation teams must document the provenance of every data source. A clear audit trail helps demonstrate compliance to regulators or customers.
Risk mitigation in production: Legal fallout could lead to penalties or shutdowns for services that depend on scraped data. Building redundancy—using multiple licensed sources—reduces this risk.

FAQ

Q: Should I stop using scraped data in my workflows?

A: If you’re unsure whether the data is licensed, replace it with openly licensed or proprietary data. The tightening legal environment makes uncertainty costly.

Q: How can I verify the licensing status of a dataset?

A: Check the dataset’s metadata, contact the provider directly, or use tools that flag licensing terms. For web‑scraped content, look for explicit permissions or public domain status.

Q: Will this affect open‑source models like GPT‑4o or Claude?

A: Open‑source models that rely on publicly available corpora are less likely to be affected. However, if the training data includes scraped content, the same scrutiny could apply. Stay informed about the training pipelines of the models you use.

Originally published on Automations Cookbook.

OpenSpec: Lightweight AI Spec Framework for Production Workflows

Felipe L — Thu, 17 Sep 2026 22:45:48 +0000

What Happened

OpenSpec is a lightweight framework for defining AI workflow specifications. Its declarative syntax lets developers describe inputs, outputs, and validation rules for AI agents and automation flows. The platform adds tooling for versioning specs, generating documentation, and integrating with CI/CD pipelines.

The release targets teams that already use n8n or custom AI‑agent orchestration stacks. By adding a formal spec layer, OpenSpec catches configuration errors early, enforces data contracts, and simplifies auditing and maintenance of complex AI pipelines.

Why This Matters for Builders

Rapid Prototyping: A concise spec language eliminates boilerplate validation code, speeding feature cycles and lightening developer cognitive load.
Improved Reliability: Spec validation runs automatically in CI pipelines, catching mismatched inputs or missing fields before production. Runtime failures drop and user trust rises.
Clear Documentation: OpenSpec turns spec files into human‑readable docs. Ops teams can reference them during incident response, and new hires onboard faster.
Version Control & Rollback: Every spec version lives in source control, so rolling back to a previous contract is trivial if an AI model update breaks downstream consumers.
Toolchain Compatibility: The framework ships adapters for popular workflow engines (n8n, Airflow, Prefect) and fits into existing Docker or Kubernetes deployments.

FAQ

Q: Does OpenSpec replace existing validation libraries?

A: No. OpenSpec complements libraries like Pydantic or Marshmallow, focusing on declarative specs while those libraries handle runtime validation.

Q: How easy is it to adopt in an existing n8n workflow?

A: Add a spec file next to your n8n JSON, run the OpenSpec CLI to validate, and hook the validation step into your CI pipeline. No changes to the n8n runtime are required.

Q: Can I use OpenSpec with multiple AI models in one workflow?

A: Yes. Each model or node can have its own spec, and you can reference shared schemas across them, keeping contracts consistent throughout the pipeline.

Originally published on Automations Cookbook.

Skillsync Launches Portable AI Chat Sessions for Agents

Felipe L — Thu, 17 Sep 2026 18:51:47 +0000

What Happened

Skillsync, a YC W26 startup, launched on Hacker News. The company offers a lightweight API that lets AI chat sessions move between agents without losing context. A demo shows a user’s chat passing from a front‑end bot to a back‑end analytics agent while keeping the conversation state intact. The team says the solution is agent‑agnostic and works with OpenAI, Anthropic, Llama, and other frameworks.

Why This Matters for Builders

Simplified state management: Builders can use Skillsync as a single source of truth instead of each agent keeping its own session store, cutting boilerplate and state drift.
Cross‑agent collaboration: Multiple specialized agents—such as a scheduling bot and a data‑analysis bot—can share conversational context, producing richer interactions.
Reduced latency: Offloading session persistence to a dedicated service lets agents focus on computation, which can lower response times.
Easier compliance: A central session store simplifies auditing and enforcing data‑handling policies across all agents.
Rapid prototyping: Developers can experiment with multi‑agent conversations without writing custom persistence layers, speeding up iteration.

FAQ

Q: Does Skillsync replace my existing database for session storage?

A: No, it sits on top of or alongside your database, offering a unified API for session access.

Q: Is the API compatible with n8n and other workflow tools?

A: Yes, Skillsync provides a RESTful interface that any workflow engine—including n8n, Zapier, and custom scripts—can call.

Q: What happens to privacy‑sensitive data in portable sessions?

A: Skillsync supports fine‑grained access controls, letting you restrict which agents can read or modify specific conversation segments.

Originally published on Automations Cookbook.

PS5 Linux Lead Leaves Over LLM Misuse

Felipe L — Thu, 17 Sep 2026 12:51:50 +0000

What Happened

The PS5 Linux project lead resigned after a surge of LLM‑driven contributions. He called the current state “a bunch of noobs using LLMs that they don’t even understand.” Core maintainers warned that shallow contributions add technical debt and potential security flaws. The exit signals a crossroads: the community must choose between more LLM code or stricter review.

Why This Matters for Builders

Quality Assurance in Automation Workflows

LLMs can slip in subtle bugs. A typo in an endpoint or wrong data transform can bring a critical service down.
Security Implications

Generated snippets may include hard‑coded credentials, insecure API calls, or outdated dependencies. These can be exploited in production.
Maintainability and Technical Debt

Poorly documented, generic LLM output makes debugging hard. Future developers will pay more to update or add features.
Governance and Compliance

Regulated industries need audit trails. LLM code without provenance can breach compliance standards.

The resignation highlights the need for strict code review, solid documentation, and thorough testing when adding AI‑generated parts to production automation.

FAQ

Q: How can we ensure LLM‑generated code is reliable in our workflows?

A: Use a three‑step review: run automated linters and unit tests; have a human check logic and security; keep clear documentation and version control.

Q: Should we avoid using LLMs for core workflow logic altogether?

A: Not necessarily. LLMs speed prototyping, but critical paths—authentication, data validation, error handling—should be hand‑crafted or come from vetted libraries.

Q: What tools can help guard against LLM‑related risks?

A: Static analysis, dependency‑scan utilities, and CI pipelines that enforce standards. Add AI‑specific linters to flag unsafe patterns in generated code.

Originally published on Automations Cookbook.