Northflank is the ninth AI platform breached in October. I read its status page like a timeline.
Northflank's status page has one sentence every platform dreads writing, and it wrote it in plain language: "Assume all environment variables have been decrypted." No hedging, no "out of an abundance of caution." If you stored secrets in Northflank, the company told you on October 30 to act as if they were all gone — workloads, secret groups, global secrets — and to rotate everything. I read the whole status page as a timeline, because the gaps between updates tell you what the incident actually cost in time.
The timeline, with the gaps counted
All times UTC, from Northflank's own status page (October 30–31):
Oct 30 11:43 Wiz sends a threat alert
Oct 30 11:44 Northflank engineers review the alert (+1 min)
Oct 30 12:02 initial network policy patch (+18 min)
Oct 30 12:50 additional patching deployed (+48 min)
Oct 30 13:00 Wiz incident-response (CIRT) support requested (+10 min)
Oct 30 13:05 additional patching deployed (+5 min)
Oct 30 13:40 first customer warning: rotate env vars (+35 min)
Oct 30 13:58 "Assume all environment variables have been
decrypted. Disable/rotate all keys." (+18 min)
Oct 30 15:12 secret rotation enabled for Postgres and MongoDB;
"multiple users are confirming that their AI
keys have been used" (+1h 14min)
Oct 30 19:28 "investigation is ongoing... no further
confirmed findings" (+4h 16min)
Oct 30 23:38 users report unauthorized use of AI API keys;
rotate at the issuing provider, check provider
usage and billing logs (+4h 10min)
Oct 31 01:04 audit logs enabled for all customers,
including free users (+1h 26min)
Oct 31 13:34 secret rotation for MySQL and Redis now
available to all customers (+12h 30min)
Two numbers stand out. First: 1 hour 57 minutes from the Wiz alert to the first customer-facing warning, and 2 hours 15 minutes to the blunt "assume everything is decrypted" version. That is fast for an incident response, and it still wasn't fast enough — by 15:12, barely three and a half hours after the alert, users were already confirming their AI keys had been used. Second: 12 hours 30 minutes between audit logs for everyone and rotation tooling for MySQL and Redis. Rotation tooling rolls out database by database, which means the remediation itself was a queue, and the last customers in that queue waited more than a day from the alert (25 hours 51 minutes, alert to MySQL/Redis rotation).
Why "assume all decrypted" is the honest sentence
The status page never says what was taken, how the attacker got in, or how many customers are affected. As of the latest update I read, there is no published root cause and no statement about whether bring-your-own-cloud customers are affected. In that vacuum, "assume all decrypted" is the only instruction that doesn't depend on facts the company doesn't have yet. The alternative — "rotate the keys we think were touched" — would have been a guess dressed as precision.
It also matches the pattern of the whole month. Wiz's "Black October" research, published October 30, counts nine AI developer platforms disclosing incidents in October alone:
Oct 6 Fireworks AI
Oct 9 Andromeda
Oct 14 Zeabur
Oct 15 Lightning AI
Oct 16-17 OpenCode
Oct 20 Railway
Oct 23 Warp
Oct 28 CommandCode
Oct 30 Northflank
Wiz says there is no single vulnerability or dependency linking them. What links them is the target: stored environment variables, with AI keys first in line. And Wiz traced what happened to two enterprise Bedrock keys stolen in this wave: tested within 30 minutes of the reported credential access — a single model call from a residential IP, part of a one-minute sweep across many organizations — then left nearly silent for four and seven days, then roughly 2.5 million calls and over $100,000 at list price. The quiet days are the part that should scare you: a stolen key that passes a spot check on day one can still detonate on day seven. Northflank's advice to check provider billing logs, not just rotate, is aimed exactly at that window.
Two things to keep straight. Wiz does not say the two traced Bedrock keys came from Northflank — only from "one of these vendor breaches." And nobody has shown the nine incidents share one attacker. The shared fact is narrower and worse: a platform that stores its customers' secrets in one place only needs one break-in to pay off many times over, and this month it happened nine times.
What to actually do, in order
If you have anything in Northflank, the status page's own ordering is right: revoke at the issuing provider first (rotating inside Northflank alone leaves the old key valid wherever it was copied to), then AI keys and payment keys, then database and cloud credentials, then everything else — and check provider usage logs for the quiet window, not just the loud one. If you don't use Northflank, the month still applies to you: any AI key sitting in a third-party platform should be treated as a credential with a resale market, because this month proved there is one, with same-day liquidity and a one-minute test sweep.
The uncomfortable question Northflank's timeline leaves open is the one no status page can answer yet: how does a platform decrypt — or lose control of — its customers' environment variables at all? Until that answer exists, "assume all decrypted" is not just Northflank's incident response. It's the threat model for every platform that holds your secrets.
Sources: Northflank status page, "Security incident" (updates October 30–31, 2026), read October 31, 2026; Wiz "Black October" research (October 30, 2026) as reported in contemporaneous coverage, including the nine-platform list and the two traced Bedrock keys (tested within 30 minutes, quiet 4 and 7 days, ~2.5M calls, >$100k at list price). Northflank has not published a root cause or an affected-customer count as of the updates I read; the Bedrock keys are not attributed to Northflank specifically.
Top comments (0)