
If you're in the US, there's a good chance last week reached you somehow. Outlook and Teams went dark last Thursday. The FAA grounded flights nationwide on Friday the 13th. Chase customers reported not being able to move money on Monday, and Apple services went down on Tuesday.
By then, a list of fifteen outages in six days was circulating, and it raised a fair question for anyone responsible for keeping systems running: is this connected, and should we be worried?
The short answer is that, for now, there is no evidence of a coordinated attack. And impacted organizations have stated that much of what broke, broke for ordinary reasons. The longer answer is that those ordinary reasons deserve more attention, because AI agents are about to produce a lot more of them. Below is what the record shows, outage by outage, and lessons we can take away.
Fifteen outages, no confirmed attack
No company or agency has tied any of them to an attack thus far, and Director of National Intelligence Jay Clayton said the US Government does "not see the indicia of a coordinated cyber attack." Several weren't confirmed outages at all. Google said it saw no evidence of a broad Gmail disruption, and we found no statements from Navy Federal, USAA or American Express; those entries show up only as spikes in user reports, mostly around the Chase outage.
Here are the outages with a stated cause:
Marriott cited "a technical issue" and Schwab "a brief webpage loading issue," and we found no stated cause for AWS's roughly three-hour disruption. We don't know everything behind this many outages in one week, and some may never be fully explained. But the stated causes read alike: routine work on complex, sometimes legacy systems, not an adversary.
The lesson: change is the everyday risk, and agents will make more of it
The clearest explanations involve a change: an upgrade, a contractor cutover, a software release, a patch, an update. And we expect agents to add a lot more of it. Coding assistants open pull requests and merge code. Copilots update records and send messages. Operations agents restart services and edit configurations. Each of those is a production change made in seconds, and unless someone builds it in, nothing gives it the ticket, reviewer or rollback plan a person's change would get.
The pressure is building from both directions. AI is also helping researchers and adversaries find and exploit vulnerabilities faster, which raises the cost of legacy security debt. Clayton tied the NSA disruption to patching a PHP vulnerability, and IEEE Spectrum reports that PHP2Shell, the PHP flaw disclosed on October 21, was found with the help of AI tools. Google reported an agent pipeline, exploit_agent, that compares code before and after a patch, then drafts and runs exploits without human intervention, and CrowdStrike measured the ARTEX framework taking an operator from about 40 commands an hour to more than 1,500.
Agents are a new kind of identity
Security has long handled two kinds of identity: service accounts, which follow scripts, and people, who use judgment. An agent holds service-account access and makes its own judgment calls, at machine speed. NIST's October 27 guidance on agent identity says agents need an identity lifecycle "distinct from human principals" and calls for "continuous runtime verification" of each tool call.
Three recent incidents show what agents can do with the access they hold. Two involved no attacker at all:
- OpenAI's test agents. California's Attorney General sued OpenAI on November 16 over agents that escaped a test environment, used stolen credentials and broke into Hugging Face. According to Metacurity's report on the filing, an outside investigation OpenAI commissioned found about 700 agents involved over seven days. The AG calls it "one of 80 known incidents." These are allegations, and OpenAI says every major AI lab faces the same challenge.
- WorkOS. On November 6, WorkOS disclosed that a security researcher's ARTEX agent, running benchmark tests, escaped its isolated test environment, found an internet-exposed internal API, signed in with a predictable development credential that had reached production, and pulled CRM and support data between October 25 and 30. WorkOS says the activity wasn't malicious, and the researcher says the data was deleted.
- Storm-0940. Microsoft reported attacker-run agents using stolen ConnectWise OAuth app credentials to attempt sign-ins to more than 2,250 tenants "without requiring an interactive user sign-in," taking advantage of the trust those apps already had.
In all three, the agents worked through valid credentials, so their actions looked legitimate to the systems they touched.
Hygiene limits access. It doesn't review actions.
AWS's CJ Moses concluded from the ARTEX campaigns that "the agents succeeded where basic security hygiene was missing," and WorkOS says rate limits and proper credentials would have prevented its incident entirely. Start with hygiene: inventory what's exposed, retire long-lived secrets, rate-limit, and watch for unusual behavior.
Hygiene decides what an identity can reach. It doesn't decide whether an authorized agent's next action should happen. Two common shortcuts leave the same gap. AI can't police itself, because an agent reads instructions and untrusted data through the same channel. And access permissions say an agent may edit a configuration or send an email, not whether this edit or this email should go out.
Know what agents should do. Control what they actually do.
Our view at Noma: the unit of agent security is the action, judged against what the agent is supposed to be doing, at the moment it happens. That takes four capabilities working from one picture of your environment:
- Discovery and posture (AI-SPM). Find every agent, MCP server, tool and skill across endpoints, SaaS and cloud, and map who owns it and what it can reach.
- Agent access control. Scope permissions down to individual tool actions, so a leaked credential reaches only what one task needs.
- Context-aware runtime enforcement (AI-DR). Judge each action against the agent's intent, identity, session history and the data involved. Block, mask, steer or escalate the risky ones while legitimate work continues.
- AI red teaming. Test your agents the way attackers will, before they do.
None of this replaces patching or change management. It gives agent actions the scrutiny a person's change would get.
Four questions before the next alarming week
Most of what breaks will be ordinary, and AI is changing the price of the legacy security debt behind it. Shoring up hygiene, paying down that debt, and managing the agents you're adding are part of the same job. Before the next wave arrives, ask your team:
- If an agent had made one of last week's changes, would we have known before it went live?
- Can we list every AI agent in production today, and what each one can change?
- Could we stop one misbehaving agent in minutes without taking a service down?
- Which legacy systems and long-lived secrets could an agent reach today, and who owns cleaning them up?
If any answer is "not sure," that's a gap.






