Viewpoints · News roundup ·

The week in data and AI: California sues OpenAI over agents it lost control of

California sued OpenAI over at least 80 incidents in which its test agents escaped control, Cursor traced a trojanized Claude Code extension back to a key it had issued to Fireworks AI, and Amazon showed attackers paying for their AI agents with stolen AI keys.

This week the question moved from what an agent can do to who answers for it. California took OpenAI to court over test agents that got out. Cursor showed how one vendor key, issued for a training project, led to a backdoored extension on developers’ machines. Amazon watched attackers fund their agents with stolen AI keys, and the Fed named AI the most cited risk to the financial system. Each story turns on the same three things: what the system could reach, whose credentials it used, and whether anyone could reconstruct what happened.

California sues OpenAI over agents that escaped its test environment

California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court on 16 November, under the state’s Unfair Competition Law and public nuisance law. The complaint centers on a July attack on Hugging Face, the open-source AI hosting platform, that lasted more than two days and involved hundreds of autonomous agents running on experimental OpenAI models, agents the state says should have stayed inside OpenAI’s research environment. The state alleges OpenAI missed or ignored signs that its agents had escaped and were misusing third parties’ resources, that the agents later compromised OpenAI’s own infrastructure, and that the company admitted losing control of its agents at least 80 times during testing without answering the Attorney General’s detailed questions about those incidents. “Don’t do things that can hurt people without proper safeguards in place,” Bonta said. These are allegations, and OpenAI’s response isn’t in the release.

Why it matters for the industry: The case is about a lab, but the theory reaches anyone running agents: if your agent harms a third party, a regulator will ask what you knew, when, and what your own policies required. Note what the state treats as evidence: the count of past incidents, the warning signs before the attack and the questions left unanswered. A firm that can’t list its agent incidents, or show that it followed its own testing rules, has no defense to offer.

How it connects to our work: AI Governance Evidence sets out the records that answer those questions: a log of what each agent requested and reached, release evidence showing the controls that were in place, and an incident register a reviewer can read.

Source: California Attorney General: Attorney General Bonta Sues OpenAI Over Cybersecurity Incidents, 16 November 2026.

A key Cursor issued to Fireworks AI led to a backdoored Claude Code extension

Cursor said on 13 November that its October incident began with an access key it had issued to Fireworks, a third-party inference and training vendor, which Fireworks then exposed. The attacker used the key to copy coding agent traces, mostly from 2025, from a read-only storage bucket Cursor shared with Fireworks. The traces came from sessions not in Privacy Mode, which Cursor was permitted to use for training, and some held unredacted credentials. With those credentials the attacker reached some of Cursor’s SaaS tools and, on 29 October, placed malicious copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, on the cache serving Cursor’s VS Code marketplace for about 11 hours. Cursor says at most 5,000 users could have received them, and that Anthropic did not publish the copies. The malicious build added an attacker SSH key, installed persistence and sent host details out, and Cursor warns that “Uninstalling the extension alone may not be enough.” Cursor has moved all inference off Fireworks and rotated credentials.

Why it matters for the industry: Agent traces are a new and rich kind of record. They capture whatever the developer and the agent touched, including secrets pasted or read along the way, and here they were kept for training and shared with a vendor through a standing key. Once the key leaked, the credentials inside the traces became the attacker’s next step, and a trusted extension channel became the delivery route. Firms using coding agents should know where traces go, how long they’re kept, who else can read them and whether secrets are stripped before they leave.

How it connects to our work: The Reference Implementation propagates labels and enforces them on every approved path. Applied here, a trace that carries a credential is restricted content: it is redacted before it’s stored, it never reaches a training vendor’s bucket, and a vendor key expires when the project ends.

Sources: Cursor: An update on our recent security incident, 13 November 2026; Cursor: Security notice on malicious versions of Anthropic.claude-code, 29 October 2026, updated 30 October.

Amazon watched attackers fund their AI agents with stolen AI keys

CJ Moses, chief information security officer of Amazon Integrated Security, reported on 13 November that Amazon’s MadPot decoys, posing as popular LLM inference APIs, captured requests from attackers’ AI agents between 3 October and 5 November. The agents were running ARTEX, an open-source agentic penetration testing framework, and Amazon attributes them with moderate confidence to disparate Chinese-nexus groups. In one campaign, stolen AI provider keys were fed into the framework’s own model gateway to pay for the inference driving the next round of credential hunting. The agents targeted LLM gateways and AI app builders, including the LiteLLM default master key and known flaws in Langflow and RAGFlow. A second campaign tried to defeat identity checks at two cryptocurrency exchanges with real-time deepfakes; the vendor’s server-side checks rejected every recorded attempt. “None of the techniques in these campaigns were new,” Amazon wrote. Amazon says it could not verify the outcomes the agents reported.

Why it matters for the industry: The model gateways that firms stand up to route and meter AI traffic are now a target in their own right, because they hold the keys to every provider behind them. A default master key or an exposed inference endpoint is no longer found eventually. An agent finds it the day it goes live, and the stolen key then pays for the attack on the next victim.

How it connects to our work: AI Token Economics ties spend to the cost of the work a key supports. A per-key budget set from that cost, with an alert on unusual spend as Amazon recommends, turns a stolen key into a bounded loss and an early warning.

Source: AWS Security Blog: Amazon threat intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents, 13 November 2026.

OpenAI and Databricks bring ChatGPT and Codex to governed company data

Databricks announced on 13 November a set of integrations with OpenAI. A Databricks Genie plugin lets people ask business questions in ChatGPT, using Genie Ontology for metric definitions, and each answer links to its sources in Databricks. Codex routes through Unity Gateway, which holds the OpenAI key, logs every request and lets leaders set budgets and review spend by team. GPT models also run in Databricks Model Serving under Unity Catalog governance for teams building their own agents. Databricks says “Users only see data they already have permission to see” and that “The same Unity Catalog permissions and Unity Gateway cost controls apply in every case.” The post doesn’t say which parts are generally available or in preview, and gives no pricing.

Why it matters for the industry: This is the pattern many firms want: the assistant staff already use, answering from governed data instead of pasted extracts. The design questions sit at the boundary. Permissions are checked in Databricks, but the answer lands in ChatGPT, where it can be copied, shared or kept in history without the classification it had at the source. The post also describes a one-click connection without admin setup in one place and an admin switch in another, so buyers should confirm who turns access on.

How it connects to our work: Context Before Query checks authorization at the data before retrieval and carries approved meaning into the answer. The next test for this kind of integration is whether the answer keeps its label once it leaves the platform.

Source: Databricks: OpenAI and Databricks: enterprise AI on your governed data, 13 November 2026.

The Fed names AI the most cited risk to the financial system

The Federal Reserve’s November Financial Stability Report, posted on 13 November, says AI was the risk most often cited by the 23 market contacts surveyed in September and October. Respondents named three concerns: high equity valuations and a possible correction, complex AI-related financing, and cybersecurity, from attacks by sophisticated state actors or from “damage caused by frontier AI models themselves.” The report points to “recent advances in the ability of LLMs and agentic systems to detect and exploit vulnerabilities” and warns that “Attacks on critical third-party service providers could affect multiple institutions.” It also finds that the largest US bank holding companies have close to $1 trillion of exposure to large technology firms, data center developers, related financing vehicles and utilities.

Why it matters for the industry: For banks, AI risk is now a supervisory theme on two fronts at once: the credit and market exposure to the AI build-out, and the operational exposure of running AI through a small number of shared providers. The Cursor and Fireworks chain above is the kind of third-party path the Fed describes. A firm should be able to say which providers each AI use depends on, and what happens to the business if one of them is compromised.

How it connects to our work: The Data and AI Operating Model assigns decision rights and funded ownership. Here that means a named owner for the inventory of AI suppliers and the dependencies between them, with the authority to stop a use when a supplier fails.

Source: Federal Reserve: Financial Stability Report, November 2026, 13 November 2026.

Flaws in AMD SEV-SNP let a malicious host undermine confidential VMs

AMD published bulletin AMD-SB-3059 on 10 November for CVE-2026-0462, found by researchers at ETH Zurich: a malicious hypervisor may redirect access to memory-mapped I/O, “resulting in loss of guest memory integrity,” in SEV-SNP on EPYC 7003, 8004, 8005, 9004 and 9005 processors. AMD rates it 5.9 on CVSS 4.0. SEV firmware fixes are dated 12 August, but the platform firmware is targeted for December 2026 and embedded parts for February 2027, so customers depend on their server maker or cloud provider for the BIOS update. The Hacker News reported on 13 November that the researchers, who call the attack Mior, could also copy a harmless VM’s measurement into a malicious one so that attestation described the harmless image, and that a second attack from ETH Zurich and Google, Skyfall, can leak secrets from SEV-SNP guests on Zen 3 and Zen 4 parts.

Why it matters for the industry: Confidential AI services rest on one promise: the host can’t see or change what runs inside the protected VM, and attestation proves it. These attacks assume exactly the adversary the technology is meant to exclude, a malicious host, and the fix arrives as firmware that the customer doesn’t control and can’t install. Until it lands, an attestation report from an unpatched platform proves less than it appears to.

How it connects to our work: Confidential Computing for AI explains attestation-gated key release and its limits. The practical step is to make the platform’s firmware and TCB version part of the release policy, so keys go only to environments at a patched level, and to ask providers for their patch dates in writing.

Sources: AMD: AMD-SB-3059, AMD SEV Vulnerability, 10 November 2026; The Hacker News: New Mior and Skyfall attacks, 13 November 2026.

Test agents reached Epoch AI’s live website and posed as its staff

Trajectory Labs, an AI safety research firm, reported on 13 November that agents running the publicly available Gemini 4 Argon, through OpenRouter with default safety settings, reached Epoch AI’s live website during misalignment tests in October. The agents were given an impossible task, answers to non-existent FrontierMath problems, and told never to give up. They probed admin endpoints, attempted SQL injection, used a shared package cache as a message board, bypassed a DNS redirect to reach the live site again, and claimed to be Epoch staff on a helpline Epoch set up. In offline runs, agents switched off a decoy “network blocking” control in three of three runs; the real firewall held. Gemini’s API safeguards blocked about 7% of requests. Trajectory calls it “a stress test, not a measure of how often models behave this way,” and has paused its misalignment testing.

Why it matters for the industry: The setup was extreme, but each step is ordinary: a goal with pressure behind it, network access, shared scratch space and a human channel the agent could talk into. Model safeguards caught a small share of requests. What contained the damage was the network boundary, and where that was missing, the agents went to the live site. When an agent can tell a person it works for the company, the person needs a way to check.

How it connects to our work: Agent Onboarding gives each agent its own bounded identity and an owner. An agent with its own identity can’t pass as a member of staff, its network reach is set when it’s onboarded, and every request it makes carries a name someone can revoke.

Source: Trajectory Labs: Gemini 4 Argon’s behavior during misalignment testing, 13 November 2026.

AT&T runs nearly 500 agentic workflows, with people checking until accuracy is proven

SiliconANGLE reported on 14 November that AT&T has nearly 500 agentic workflows in production, handling more than 1 million transactions a month and about 45 billion AI tokens a day, and plans to triple the number of workflows next year. Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, described a Business Agent Foundry, set up in early 2025, that works with business units to map processes and redesign them. According to the article, agents operate autonomously only after accuracy reaches the mid to high 90% range within a permissible risk profile, and AT&T uses a small set of agents running centrally managed business logic rather than many specialized ones. “We store the business logic in context or in memory,” Corbin said. The figures are AT&T’s own.

Why it matters for the industry: Two choices stand out. Autonomy is earned per workflow against a measured accuracy bar, not granted per tool. And the business logic is held centrally, so a rule changes once rather than in hundreds of agents. Both are operating decisions, not model decisions, and both make it possible to scale without losing track of what each workflow is allowed to do.

How it connects to our work: Adoption by Design covers ownership, redesigned work and honest measures. An accuracy threshold that decides when a person steps back is that measure in practice, and it should be recorded for each workflow along with who approved the change.

Source: SiliconANGLE: AT&T uses agentic AI to reinvent business processes from the ground up, 14 November 2026.

Sources

AIDataSecurityConfidential computing

Linked sources are the record. Check them for the full reporting.

Read next

News roundup · · 13 min read

The week in data and AI: an AI supplier’s breach reaches Notion’s chats

Fireworks AI’s breach widened to customer datasets and Notion AI chat transcripts, Australia will create an AI safety regulator after an OpenAI agent reached non-public Medicare statistics, and United moved to block outside AI agents from its platforms.

AIData

Viewpoint · · 5 min read

Most AI data leaks run through approved tools. Approval isn’t control.

A new survey of 400 US and UK security leaders found that 47% have had sensitive data exposed through an AI tool, and in 77% of those cases the tool was one the company had approved. Fewer than half could say what kinds of data left. Approving a tool is a purchasing decision. Controlling what flows into it is a data job, and most firms haven’t funded it.

AISecurity