A month that earned its nickname
Between 6 and 30 October 2026 the AI toolchain stopped looking like infrastructure and started looking like a target range. At least nine AI developer platforms disclosed security incidents inside 25 days:
- Fireworks AI (6 Oct)
- Andromeda (9 Oct)
- Zeabur (14 Oct)
- Lightning AI (15 Oct)
- OpenCode (16–17 Oct)
- Railway (20 Oct)
- Warp (23 Oct)
- CommandCode (28 Oct)
- Northflank (30 Oct)
Cloud security firm Wiz published the tally on 30 October and the industry promptly named the month “Black October”. The list has kept growing: NVIDIA's Brev, Trigger.dev (1 Nov) and Chainstack were added days later. Any other month, one of these would have been the security story of the week.
What was actually taken
There is no single zero-day behind this. Wiz is explicit that no shared vulnerability or dependency links the nine incidents, and the reported root causes differ: credential access at Fireworks, Zeabur and Lightning; “privilege escalation in our build infrastructure” at Railway; a third-party vendor dependency at Warp; a “legacy system” at OpenCode.
What every report does share is the target: attackers hunted stored environment variables, above all the ones that unlock AI models. Fireworks, OpenCode and Warp reported indications of an “agent swarm” attack; Andromeda called its incident part of an “industry-wide attack”; OpenCode and CommandCode attributed theirs to an attempt at distillation. Wiz ties the whole picture to two trends it has tracked all year — agentic attacks that let intruders move at machine speed, and an LLMjacking economy whose “LLM brokers” make a stolen key instantly monetisable.
From stolen key to token flood
The most concrete evidence comes from two enterprise keys that Wiz could follow end to end. In both cases the first unauthorised activity came within 30 minutes of the reported credential access at the vendor: a single InvokeModel call from a residential IP, part of a one-minute sweep testing stolen keys across several organisations. The call asked for Claude Fable 5.1, failed, and fell back to Opus 5.5, which worked.
- Case A — the permissive key. Its policy allowed
global.*cross-region inference profiles. Over thirty hours the actor made about 800,000 Bedrock calls, roughly 90% successful; global routing spread the load so only about 6% were throttled. Opus models were three quarters of the successful traffic. - Case B — the restricted key. Limited to
us.*profiles, it surfaced in a reseller pool instead. One Level 3 address still managed about 1.7 million calls in thirteen hours, over a million of them successful, the rest capped by regional quotas.
At list price, those two keys alone likely ran up over $100,000 in usage. The token profile — prompt-cached inputs, a long tail of heavy calls — looks exactly like resold coding-agent traffic, and Wiz traced the same keys through overt resale hubs selling discounted frontier models.
Why this one is different
The impact was cheap to produce and easy to sell. Public reports mention roughly $371,000 in cryptocurrency allegedly stolen via keys stored on Railway, and consumer bills in the hundreds to thousands of dollars from abused AI access. And the collateral is structural, not just financial: Zeabur published a full write-up and is shutting down because of its incident.
That is the uncomfortable part of the pattern. A restrictive key policy did not prevent the theft, but it did force the attacker to spend hours finding the permitted profiles and left quotas as a real ceiling — the difference between 800,000 and 1.7 million calls landing on your bill.
What follows from it
For anyone running coding agents, the lesson is blunt: the provider key your agent uses is exactly the credential these attackers were hunting.
- Rotate first, ask later. If keys lived on any of the named platforms, treat them as exposed — rotate everything, not just the AI keys.
- Narrow the policy. Restrict inference profiles to the regions and models you actually use; “allow everything” is what Case A paid for.
- Set budgets and alarms on invocations and spend per key; the flood arrived within hours of the first validation call.
- Prefer short-lived, role-based credentials so a stolen key is a smaller prize.
The month also makes a broader point about hype and evidence. For once the alarming headline survives scrutiny: nine named companies, dated disclosures, published timelines and telemetry that shows real money leaving real accounts. The boring mitigations — rotation, narrow scopes, quotas, budgets — are what actually moved the needle here, not another model launch.
Sources:
- Wiz, “Black October” for AI Developer Platforms: https://www.wiz.io/blog/black-october-for-ai-developer-platforms
- Fortune on the “Black October” hacks: https://fortune.com/2026/10/29/black-october-ai-cyberattacks-hackers-php2shell
- Zscaler ThreatLabz on the AI-assisted ARTEX campaign: https://zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
