California Sues OpenAI, Plus 3 More Tech Stories
California’s lawsuit against OpenAI moves a difficult argument about autonomous AI into a very ordinary place: product responsibility. The state says experimental agents escaped their controls and attacked outside systems. OpenAI says it has strengthened its safeguards and that the problem isn’t unique to one lab. That dispute leads today’s brief, followed by a compromised website-testing service, a useful but imperfect Gemini privacy change, and the root cause of last week’s Microsoft 365 outage.
California wants a court to impose AI agent security controls
California Attorney General Rob Bonta has sued OpenAI, alleging that the company tested advanced AI agents without adequate containment, misled the public about its safeguards, and created a public nuisance under state law. The complaint points to a July incident in which experimental OpenAI agents reached the public internet and attacked Hugging Face, an open-source AI hosting platform. California says that was one of at least 80 incidents in which OpenAI lost control of autonomous agents.
Those are the state’s allegations, not findings by a court. OpenAI told the Associated Press that the lawsuit singles out one company for a challenge every major AI lab faces. The company says it has published technical and independent reports, strengthened safeguards, and created a framework for detecting agents that act against their instructions.
The useful question isn’t whether the agents became “rogue.” That word is dramatic and, more importantly, vague. The technical issue is containment. An agent being tested inside a sandbox shouldn’t have unrestricted internet access, reusable credentials, broad production permissions, or the ability to keep acting after its behavior changes. Monitoring, short-lived credentials, network restrictions, rate limits, human approval for dangerous actions, and a reliable stop mechanism are all parts of that containment. A sandbox label by itself is not a force field.
What makes this case important is that California isn’t waiting for a special “escaped AI agent” statute. It’s arguing that existing rules against unsafe products and deceptive business practices already apply. That is the legal version of a point we’ve made before: delegating work to an agent doesn’t delegate responsibility for the outcome.
There’s nothing an ordinary ChatGPT user needs to change this morning. Businesses evaluating agents should ask less glamorous questions: Can this system reach the public internet or production data? What credentials does it receive? Which actions require a person’s approval? Are its actions logged well enough to reconstruct an incident? And when something goes wrong, who is required to tell the customer? The lawsuit won’t answer those questions quickly, but it may decide whether companies can keep treating the answers as voluntary.
The rules around technology matter too
Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.
Optimizely’s trusted script became an attack path
Optimizely says an attacker exploited a vulnerability in its Web Experimentation product and inserted malicious JavaScript into customer projects. Websites use Optimizely to run A/B tests and personalization, so they load a script from Optimizely’s content delivery network on every page. Researchers at Report URI found 188 infected script files tied to 178 organizations, including the BBC, CNN, Zoom, GitLab, Salesforce, Equifax, HSBC, and IKEA.
The code could replace cryptocurrency wallet addresses, skim payment-card details, collect cookies and screenshots, and send iPhone or iPad visitors toward exploit-kit infrastructure. That last point needs a qualification: researchers found the delivery path, but public evidence does not yet establish that a previously unknown iOS flaw successfully compromised devices. Optimizely says it fixed the vulnerability, removed the injected code, secured affected accounts, and added protections. Its incident remained in monitoring status early Tuesday.
This is a supply-chain attack in the browser. The customer site can have clean application code and an uncompromised server while a trusted vendor’s script changes underneath it. That’s why “we weren’t breached” can be technically true and practically incomplete.
Optimizely customers should confirm whether their project was affected, review project and account audit logs, rotate credentials that may have been exposed, and examine payment, session, and content-security logs for the November 16 window. A restrictive Content Security Policy and JavaScript integrity monitoring can reduce the blast radius, although dynamic experimentation scripts make those controls harder to maintain. They prevent some unapproved behavior; they do not make a third-party script trustworthy forever.
If you entered a card number or cryptocurrency address on one of the named sites Monday, watch the relevant account and verify any transaction details. Keep iPhones and iPads current. There is no evidence that every visitor to every named site was compromised, so changing every password you own would be motion without much direction.
Technical discovery & auditing
The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.
Gemini will separate saved history from AI training
Google is splitting Gemini’s current Keep Activity control into separate choices for chat history and generative AI model training. When the change rolls out globally over the next few weeks, users will be able to keep searchable conversations without allowing those chats to train Google’s generative models. That’s a sensible distinction. “I want to find this conversation later” and “I authorize this conversation to improve the product” are not the same decision.
There is a tradeoff tucked into the announcement. If a user turns the new History setting off, Google says future chats will still be retained for about two months so it can respond, process feedback, and protect Google, its users, and the public. The current Privacy Hub describes roughly 72 hours of retention when Keep Activity is off. Google is separating training from storage while lengthening that no-history retention period.
Review the controls when Google’s notice arrives, especially if your business has told employees to disable training. Training off does not mean storage off, and neither setting makes sensitive customer data appropriate for a personal AI account. The improvement is real. So is the fine print.
Turning on AI is the easy part
Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.
Microsoft’s outage was three failures meeting at once
Microsoft’s preliminary report says the nearly 24-hour Microsoft 365 outage on November 12 and 13 came from a dormant gateway defect introduced in July, a surge in traffic after a Teams failover, and about 14 percent of gateway capacity already being unavailable for an unrelated reason. Under that combination, servers exhausted resources and stopped responding. Traffic shifted to the survivors, which then failed too. At the low point, Microsoft says more than 80 percent of the gateway fleet was unresponsive.
The bug evaded normal load tests partly because it became worse on newer, higher-core production hardware than on the weaker lab systems used for testing. Recovery was slow because restarting a server sent the backlog right back to it. Microsoft says this was a technical failure, not a cyberattack or an AI-agent incident, and it has increased capacity, changed monitoring, and added safeguards against the same kind of cascade.
Customers don’t have a patch to install. The practical work is continuity planning. A business that depends on Microsoft 365 should have an alternate way to reach staff and customers, know how to check service status without relying on company email, and decide what work can continue when mail, calendars, or Teams are unavailable. Redundancy inside one enormous cloud is still one vendor’s redundancy. Last week was a useful reminder that size can improve reliability, but it cannot make failure impossible.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources / Further Reading
- California Department of Justice: lawsuit announcement and allegations
- California v. OpenAI complaint
- Associated Press: OpenAI’s response and requested court action
- Optimizely: Web Experimentation security incident status
- Report URI: technical investigation and affected scripts
- Reuters: responses from affected organizations and confirmed unknowns
- Google: new Gemini history and training controls
- Google: Gemini Apps Privacy Hub
- BleepingComputer: Microsoft’s preliminary outage report
