Two million prompts went out, each carrying a secret nobody had ever used. Some of those secrets came back, in the hands of strangers.
Which is a pretty expensive way to learn that "private" is a setting, not a guarantee.
This is Lantern Cybersecurity Daily, with a Saturday edition for Saturday, November 21. Coming up: Bell Canada says an "agentic" attacker took enterprise customer data, a canary experiment that caught AI providers and resellers leaking prompts, a Cisco Expressway flaw now under attack, Amazon on AI agents that found a zero-day by themselves, and the real cause of N-able's email security outage.
Tap follow so the next episode finds you.
BleepingComputer, on a breach notice from Bell Canada:
According to the notification, an unauthorized party accessed a pre-production environment of Bell's Maximo system, which the company uses to order and provision telecommunications network connections.
Bell is one of Canada's largest telecom providers. It confirmed to BleepingComputer that on Tuesday, November 17, what it calls an "agentic threat actor" got into some of its systems, including the one used to order and provision circuits, and took enterprise customer data.
So who's affected? Business customers, not consumers. Per the notice, the exposed data includes names, job titles, corporate emails, phone numbers and business addresses, plus ticket titles and some IP addresses for network elements. Bell says there were no credentials, attachments or highly sensitive personal data in the set, and no impact on customer services.
Slow down on one word, though. "Agentic" is Bell's term, and BleepingComputer reads it as a sign that AI tools were used, but Bell hasn't explained what it saw, says it doesn't know who did it, and hasn't said how many customers were hit or how much data left.
It also says it's sharing indicators of compromise with enterprise customers and the Canadian Centre for Cyber Security, but not publicly, and it didn't answer when BleepingComputer asked for them. If other telecoms are supposed to check for the same visitor, that's not much help.
And think about what's on that list: real contacts tied to real circuit orders and network IPs. Bell's own warning is to watch for anyone impersonating Bell. For an enterprise network team, a fake ticket that cites your actual circuit is a very convincing pretext.
Researcher HD Moore, chief research officer at Dragos, quoted by Dan Goodin in Ars Technica:
Each key was sent in exactly one request to one provider, all within a few short windows in late October, and every request we sent during those windows was stolen.
The setup is clever. Moore seeded nearly two million prompts with unique bait: working AWS keys tied to empty accounts, SSH passwords, secret URLs and private git repository references. Each secret went into exactly one prompt to one provider, so if it surfaced later, he could trace which link in the chain leaked it. The targets were AI routers like OpenRouter or Vercel, resellers, and the hosting providers behind them.
He published Thursday, and there are three findings. First, he found evidence that at least three AI providers had been breached by attackers who siphoned the prompts they received. Second, some resellers offered low prices, logged the prompts, and the credentials in them got used. Third, some providers ignored zero data retention and no-collection flags.
And the stolen keys were put to work. Per Moore, attackers sometimes tested them within hours, combed S3 buckets and secrets managers for more credentials, and added their own IAM users. The bait keys were used tens of thousands of times in AWS API calls.
The one that stings is a Chinese AI lab whose systems fetched the secret URLs from more than half the prompts sent to its models, including about two thirds of the ones marked zero retention. And US providers reselling that lab's models triggered the same scraper. So if you picked a US host to keep your data in the US, it didn't.
Moore's advice runs from most to least protective: host models yourself if you can afford it, otherwise go directly to a first-tier provider, where you get one contract and one privacy policy, and choose your upstream providers instead of taking default reseller pools. He's clear that none of that guarantees nothing leaks. The cleaner fix is not to put live credentials in a prompt at all.
Eduard Kovacs, writing in SecurityWeek:
The advisory has not been revised since its initial release and still states that Cisco is unaware of public announcements or malicious use of the flaw.
Cisco Expressway is the gateway that lets remote staff and partners reach a company's video, voice and messaging systems. The bug is CVE-2026-76466, rated 9.8, which Cisco disclosed and patched on October 21. An unauthenticated attacker with network access can run commands as root, but only on clustered devices. Cisco says it comes from persistent cookies that act like fixed credentials, plus client certificates that go unchecked when TLS Verify Mode is enabled for clustering.
The evidence for exploitation comes from someone other than Cisco. Russian security firm F6 reported Thursday that, soon after a proof of concept appeared on GitHub, attackers hit a Russian organization. They broke into two internet-facing Expressway X15.5.3 boxes, then pivoted to two more cluster nodes on the internal network. On Friday, CISA added the flaw to its exploited list, with a federal deadline of Monday, November 23.
Once they had root, they staged Python scripts to scan the network, find SMB shares, talk to SQL Server and WinRM, and search GitLab for credentials. F6 says comments in the code pointed to an AI assistant, and the scripts were edited mid-attack.
If you run clustered Expressway, this justifies an interruption. The fix is X15.5.20. Anyone on X14 or earlier has to migrate to a fixed release, and Cisco lists no workaround. If you can't patch yet, F6 suggests blocking TCP ports 4371 and 4372 at the firewall except from trusted cluster nodes, and checking the Erlang log for the string F6 published as an indicator.
The bigger lesson is F6's own: one perimeter box led to the internal cluster. Patching is half of it. Segmentation decides how far an attacker gets after that.
CJ Moses, Amazon's chief information security officer, on the AWS Security Blog:
In about 11 minutes, the agents went from mapping the application’s API to arbitrarily writing files on the server.
Some background: we've been following ARTEX and other open-source agent frameworks showing up in real intrusions, and Google had already named Cairn among those in live use. Now Amazon has a detailed case. On September 10, its threat intelligence team found a Cairn instance running with no authentication, which exposed the operator's prompts, configs and the agents' output. The goal the operator gave them was simple: obtain shell access.
Over about 48 hours starting September 8, the agents went after universities in East Asia outside mainland China. Most of it failed: 16 hours and 148 actions against an open Redis server, and more than 30,000 credential combinations against a GPU server, all for nothing.
Then they hit an open-source medical imaging AI app. They found a path traversal flaw in an upload filename parameter, wrote a reverse shell into sitecustomize.py, which Python loads at startup, and triggered a training endpoint to get a root shell. Amazon assesses with high confidence that it was a zero-day, since there was no fix, issue or advisory. The maintainers have since shipped a fix after Amazon reported it.
Then the human operator handed the agents a tunnel into a second university running the same software. Amazon attributes the operator with moderate confidence to a Chinese-nexus actor, based on Simplified Chinese prompts, China Standard Time timestamps and the targeting.
So what's actually new? Not how good the agents are. Most of their attempts were ordinary and most failed. It's that they kept going around the clock, and once they found a real hole, it took them minutes. Amazon's recommendation is high-confidence detections with automated response, and humans overseeing the outcomes.
BleepingComputer, quoting N-able on its email security outage:
We are aware of unauthorized access to one of the servers in the Mail Assure and SpamExperts Hosted Cloud.
Mail Assure and SpamExperts are N-able's hosted email filtering services, used across the Americas, Asia-Pacific and Europe. On Tuesday, N-able blamed an outage on platform configuration and service dependencies. A day later it reversed that: the outage came from a security incident, and someone had gotten into a server in that hosted environment.
N-able says mail flow and spam filtering kept working, and it has no evidence so far that customer data or tenant mailboxes were accessed. But it shut off UI and API access, plus quarantine, logging and mailbox sync. Until those come back, customers may not be able to log in, release quarantined mail, check logs or add and remove domains.
N-able says this is unrelated to the N-central mess we've been tracking, where it took hosted N-central offline to patch more than 30 disclosed flaws and attackers went on to exploit one to plant cryptominers.
Unrelated, sure. It's still two separate intrusions at one vendor whose customers are mostly managed service providers. And "configuration problem" turning into "breach" a day later means the next update gets read more skeptically.
What's confirmed is one compromised server and lost management access. What's still an open question is whether anything was taken. If your mail runs through these services, don't read "no evidence" as "no access" until the investigation is done.
Two of today's stories were really about AI infrastructure, the routers and agent frameworks. If you want the builder's side of that, try AI News for Builders, by Lantern. It's top AI news for engineers, founders and investors every weekday, separating real capabilities from demo hype, fast.
Three things to do: patch clustered Expressway or firewall those ports, keep live keys out of prompts, and if you're a Bell enterprise customer, be suspicious of any ticket that seems to know your circuits.
And if a vendor says "configuration issue" on day one, check back on day two.
Lantern Cybersecurity Daily comes to you from Lantern Podcasts. We're back with our next episode on Monday, November 23.