Every AI system in your company. In one place. Under control.
Joined August 2021
- Tweets368
- Following13
- Followers617
- Likes32
A busy week across AI, from new approaches to text watermarking to some pointed questions about AI safety in New York. Here are 5 updates worth catching up on:
1. OpenAI begins text watermarking in the EU
OpenAI is rolling out textGrain, an invisible watermark for eligible ChatGPT and Codex text outputs in the EU, in response to the EU AI Act’s transparency requirements. API customers globally can opt in.
2. AI safety takes centre stage at NYC Council
Executives from OpenAI, Google, Anthropic, and Meta testified at a New York City Council hearing on AI risks. Council members pressed them on potential harms and safeguards, alongside former employees who raised concerns about the industry’s approach to safety.
3. Specialized decision models are gaining traction
Databricks and Cloudflare have introduced models designed to make structured decisions, returning choices or probabilities rather than generated text. It’s a growing area of AI infrastructure, with potential applications in classification and agent workflows where speed and cost matter.
4. Anthropic releases Claude Haiku 5.5
Anthropic’s latest small model is designed for high-volume, cost-sensitive tasks. The company says it’s faster than its previous Haiku model and costs around 75% less to run on average.
5. Anthropic invests in enterprise AI skills and ethics
Anthropic’s $100M Claude Frontier Academy aims to train 10,000 enterprise engineers by the end of 2027. The company has also brought philosophical perspectives into its AI work, inviting Hindu monk Swami Sarvapriyananda to discuss ethics and how Claude should approach moral questions.
Introducing jevals: replacing LLM judges with typed Jev decisions.
Write an eval once. Use it to test your agent and gate its tool calls.
Supports Jev, Kev and Laya. Jev is now available in Openlayer, too.
github.com/openlayer-ai/jeva…
Meta just agreed to pay up to $17.1 billion to settle claims it knowingly designed addictive features that harmed children on Instagram and Facebook. The largest tech privacy settlement in history, more than 12 times the previous record.
The case ran across COPPA, state consumer protection laws, California's False Advertising Law, and the Unfair Competition Law simultaneously. A single privacy failure drawing exposure across multiple statutes at once, with state AGs coordinating enforcement like a federal agency would.
But let’s look more closely at the settlement structure: $12 billion is paid upfront over ten years. The remaining $5 billion only releases if YouTube, TikTok and Snap implement matching restrictions and pay equivalent penalties. Meta has made clear it sees this as a moment to establish shared industry standards, calling on peers to adopt the same measures.
An independent auditor monitors compliance for five years. And thousands of individual mental health claims remain pending, meaning the financial exposure isn't close to over.
As Nicholas Baker, Head of Governance at Openlayer, put it: “While a $17.1 billion settlement headline is massive, the true impact lies in the downstream shift. Organizations across every industry should expect auditors to immediately scrutinize child privacy controls, safety frameworks, and continuous monitoring practices.”
California's Automated Decision-Making Technology regulations take effect January 1, 2027, giving consumers opt-out rights when AI makes significant decisions about them. The same enforcement machinery that produced this settlement is about to have a new set of tools.
A common miss when building test subpopulations: your filters only catch what you explicitly typed. Filtering for "billing" misses "I was charged twice" or "my invoice looks wrong," This happens constantly when people have the same intent, but say it differently.
We're excited to launch semantic search filters in Openlayer. 🔍
Instead of matching on exact words or keywords, we match by meaning. You describe the slice you care about in plain language, and Openlayer finds rows that are semantically close to that phrase, even when the wording looks nothing like it.
It's most useful for free-text columns: user messages, model outputs, transcripts, anything where the same intent surfaces in many different ways. And it composes with your other filters, so you can combine a semantic match with date ranges, user IDs, or tags to zero in on exactly the subpopulation you want to test.
Know more about it in our changelog. Link in comments!
Last September, Tractor Supply paid $1.35M to California's privacy regulator. At the time, the largest fine in CPPA history. Since then the fines have only gotten bigger:
→ GM: $12.75M for selling driving and location data from connected vehicles without telling customers
→ Disney: $2.75M for data collection and sharing practices across its digital properties
→ Total CCPA fines crossed $23M as of May 2026
The Tractor Supply case, which felt significant at the time, is now mid-table.
It started from a single consumer complaint in Placerville. The violations: no privacy policy that informed consumers of their rights, no opt-out mechanism that actually worked, and personal data being shared with third parties without contracts that contained privacy protections in place.
What made this case notable beyond the fine: job applicants were included. California extended privacy protections to applicants, employees, and contractors in 2023, meaning companies are now liable for how they handle the data of people who never even became customers. Tractor Supply failed to notify job applicants of their rights or how to exercise them.
A company with 2,500 stores selling farm equipment and workwear in 49 states. Privacy compliance wasn't their world but a single complaint in Placerville made it their problem.
A corporate officer now has to certify compliance annually for the next four years. And a California investigation can now expand to eight additional states simultaneously through the Consortium of Privacy Regulators.
As AI gets embedded deeper into how organizations at every scale handle data, automate decisions, and interact with customers, the expectation to demonstrate compliance continuously, is only going to grow.
New in Openlayer: your sessions, traces, and test results now open with an AI-generated summary. 📋
Understanding what's happening inside an AI system has always required a lot of manual reconstruction, reading through raw spans, scanning row-level results, piecing together what an agent did across a session to find a pattern.
Three surfaces now come with summaries out of the box:
- Sessions open with a summary of how the session went across its traces, what the user was trying to do, what the agent did, and where things got interesting.
- Traces show a plain-language breakdown of what happened at each step, so you're not reading spans line by line to reconstruct the sequence.
- Test results surface the dominant failure modes upfront, and you can drill down from each failure mode directly to the exact rows behind the pattern.
The goal is straightforward: understand what your AI system did and why, in natural language, without having to reconstruct it manually every time. Summaries are generated on demand and cached, so opening the same session or result again is instant!
Full changelog in the comments.
Openlayer now natively integrates with Microsoft Copilot Studio.
Conversation transcripts from Dataverse → structured traces with LLM calls, tool executions, RAG citations, and latency.
No custom logging. New agents auto-appear.
Full observability and continuous evaluation for every copilot in your org.
Docs → openlayer.com/docs/integrati…
Openlayer now natively integrates with Salesforce Agentforce.
Connect credentials → auto-discover agents → monitor production traces. No Python scripts. No custom instrumentation.
Full observability, continuous evaluation, and governance for every Agentforce agent in your org.
First native integration of its kind.
Docs → openlayer.com/docs/integrati…
AI governance isn't bureaucracy. It's what enables responsible scale.
EU AI Act high-risk obligations land Aug 2, 2026. Less than 5 months from now. The teams that are ready built the infrastructure early.
Openlayer. Connect with us. openlayer.com/contact
"We have a policy document" isn't evidence of compliance.
The EU AI Act requires provable, continuous oversight. Openlayer automates it.
Your infrastructure should be running before Aug 2nd, 2026 hits. Connect with us. openlayer.com/contact
EU AI Act high-risk obligations hit Aug 2, 2026. Less than 5 months out. And if you're a US company and your AI touches EU residents, you are in scope.
Policy sets the standard. Our technology enforces it.
Openlayer: continuous compliance mapping, real-time guardrails, audit-ready reporting.
Connect with us to start governing now. openlayer.com/contact
Openlayer partners with @TelefonicaTech to bring AI governance and observability to European and LatAm enterprises. Test. Monitor. Govern. Across the full AI lifecycle. More on the partnership: openlayer.com/blog/post/open…
New Openlayer release:
*MFA for all workspaces
*Inline image, audio & PDF traces
*Governance rule-result summaries
*Expanded SDK + API controls
AI governance must extend beyond metrics, across traces, models, and frameworks.
Full changelog → openlayer.com/changelog/stro…
We’ll be at HIMSS Global Health Conference & Exhibition, AI Pavilion, Booth 10018.
Bringing AI governance + observability to healthcare
Want to meet with our team? Set up time in advance:
Email [email protected] or DM us.
Excited to talk AI governance + observability in healthcare. See you next week in Vegas.
We’ve been recognized! Openlayer is named a Representative Vendor in the 2026 Gartner® Market Guide for AI Evaluation and Observability Platforms.
As AI adoption scales, we’re proud to provide the unified governance and observability teams need to build trust.
Check out the blog for all the details: openlayer.com/blog/post/open…
The Kolmogorov–Smirnov (KS) score compares a feature’s distribution (train vs. production) to detect drift and measure class separation.
In credit risk, fraud, and underwriting, that separation, and any shift in it, isn’t just performance. It’s risk.
Full breakdown here: openlayer.com/blog/post/ks-s…
“Needle in a haystack” failures aren’t missing data, they’re attention failures. The right context is there, but long or noisy inputs cause models to lose it.
Metrics like context utilization, relevancy, and groundedness help catch this before prod.
Read the full guide: openlayer.com/blog/post/need…
