On October 30, Wiz gave the month’s security incidents a name: Black October. By then, it was hard to argue with it.
Railway had disclosed that an attacker may have accessed customer environment variables, including some that had been decrypted. A malicious litellm release (1.107.0)had put a credential stealer on developers’ machines for roughly two hours. PHP had shipped emergency patches for two engine-level RCE vulnerabilities that researchers say had been hiding in the interpreter for more than a decade.
The name makes it sound like a single campaign. The evidence points to something messier.
Platforms that stored customers secrets were compromised, official packages shipped malware, and vulnerabilities in PHP, Node, GitLab, and a certificate authority surfaced within days of each other.
This is a snapshot as of 1 November 2026. Several platform investigations are still ongoing, so I’ve distinguished confirmed findings from preliminary reports and public claims.
The New Attack Surface Was the Platforms Holding Your Secrets
PaaS platforms typically store the credentials an application needs, from model API keys and database URLs to registry tokens, in environment variables.
The attacks weren’t aimed only at applications. Several went after the platforms sitting underneath them, where customer secrets were stored.
Wiz identified nine AI developer platforms that disclosed security incidents during October.
Fireworks
Fireworks identified unauthorized activity on October 6 and disclosed the incident three days later. A limited number of customer accounts had environment variables exposed through evaluation jobs, along with some internal credentials. Fireworks said it found no evidence that customer models, datasets, training data, or inference traffic had been accessed.
Andromeda
Andromeda said the attack likely began around October 9. It detected the incident on October 13 and disclosed it on October 20. A limited number of customer and internal credentials were exposed. The company described the incident as part of a wider attack targeting infrastructure providers.
Zeabur
Then came Zeabur on October 14–15, the incident with the most striking numbers.
Long-lived administrative credentials had been stored in plaintext and could be accessed from the public internet. The attacker used them to access the shared cluster, internal systems and some dedicated servers, eventually reading the platform database.
Zeabur’s preliminary report says the attacker accessed 3,218,965 environment-variable records, along with data associated with 321,810 accounts. The shared-cluster exposure spanned nine regions, though the company says it could not establish the full scope in every region. Separately, the attacker gained root access to 101 dedicated GCP servers and added their own SSH keys to 50 of them.
The incident was severe enough that Zeabur later announced it would wind down the company. Some services will begin shutting down on December 1, with all dedicated servers and shared clusters scheduled to stop operating by December 31.
Lightning AI
Lightning AI was hit from a different direction on October 15. Wiz linked the incident to the Mini-Shai Hulud npm supply-chain campaign. The attacker reached the platform through a compromised package and then searched for stored secrets.
OpenCode
OpenCode followed on October 16–17. This one reached users directly through a legacy Console. The incident affected 1.4% of Console users, some of whom had linked external AI provider keys to their accounts.
OpenCode described it as a large-scale distillation attack. Users subsequently reported unauthorized OpenAI and Anthropic charges ranging from hundreds to, in some cases, thousands of dollars. The CLI and desktop applications were outside the affected path.
Railway
Railway detected anomalous activity on October 20 and disclosed the incident the following day. Attackers had accessed systems storing customer environment variables. Railway said encrypted variables may have been accessed and that some may have been decrypted.
The company paused all deployments and asked customers to rotate their credentials. Its later preliminary report points to privilege escalation in its build infrastructure.
Separately, Security Alliance (SEAL) said it had triaged a report for an affected customer involving roughly $371,000 in cryptocurrency stolen across 19 EVM chains. The compromised wallet keys had been stored in Railway environment variables. Railway’s public incident report does not mention the crypto losses.
Warp
Warp became aware of an attack involving a vendor in its infrastructure on October 23. Three days later, CEO Zach Lloyd said the operation appeared to involve an agent swarm and may have started with a vulnerability in a vendor dependency.
CommandCode
CommandCode disclosed a security incident on October 28, linking it to a large-scale distillation attack targeting Anthropic and OpenAI models. It revoked all API keys as a precaution and said it had found no evidence that users were affected.
Northflank
Northflank disclosed a security incident on October 30. Later that day, it reported that users had experienced unauthorized use of AI API keys supplied to the platform. The company advised customers to assume their environment variables had been compromised and rotate their secrets. Its investigation was ongoing.
Across these incidents, attackers repeatedly targeted the platforms where developers stored application secrets.
Malware Shipped Through Official Releases
LiteLLM had already suffered a supply-chain compromise in March. On October 21, attackers reached its release pipeline through a different route.
A malicious LiteLLM release was live on PyPI for 1 hour and 44 minutes
LiteLLM is an open-source gateway that routes requests across model providers such as OpenAI, Anthropic and Gemini, while managing the credentials needed to access them.
On October 21, a malicious litellm==1.107.0 appeared on PyPI at 11:39 UTC and was removed at 13:23 UTC. Compromised Docker Hub images remained available until 14:20 UTC.
The package contained malicious Python startup hooks, including cook.pth and sitecustomize.py. These executed when Python started, even if the application never imported LiteLLM. Running pip again could trigger the malware too.
The compromise had started months earlier.
On May 27, an internal administrative key was accidentally committed to LiteLLM’s public lite-harness repository. The file was deleted three hours later, but the key was never revoked. Its original commit remained in public Git history until October 22.
The attacker used that key on October 10 to access LiteLLM’s internal agent platform, where they extracted a GitHub CLI OAuth token belonging to an organization owner. That token gave them access to the organization’s repositories and release workflows.
They used the access to steal additional maintainer credentials and modify the release pipeline. The modified workflow built from a lookalike repository, Berri-AI/litellm, instead of the official BerriAI/litellm.
The official GitHub source tag was clean. The published PyPI package and container images were not. They had been built from the attacker’s repository.
Other trusted releases shipped malware too
TensorLake
On October 8, a compromised TensorLake maintainer account was used to publish tensorlake@0.5.144 to npm. The malicious release went through the project’s own release workflow and carried valid npm provenance. Its installation hook deployed a Shai-Hulud credential stealer targeting developer machines. Only the npm package was affected; TensorLake’s other distribution channels remained unaffected.
OVHcloud CLI
On October 19, an attacker using a compromised OVHcloud developer account replaced the published binaries for ovhcloud-cli versions 0.11.0 through 0.15.0. Downloads made between 14:47 and 16:21 UTC+2 were affected.
The malicious Go binaries searched developer machines for credentials, including cloud keys, SSH material, Kubernetes configuration and AI coding-agent histories. The source code and build pipeline were untouched. OVHcloud traced the account compromise to a malicious browser extension installed on the developer’s personal MacBook in 2025.
The company removed the affected binaries and released 0.16.0 as a clean version.
PHP2Shell had been hiding in the engine for more than a decade
October 21 was also the day PHP shipped emergency security releases for two critical vulnerabilities in its core engine.
The vulnerabilities, CVE-2026-109618 and CVE-2026-109619, were grouped under the name PHP2Shell.
Both involve memory-safety problems around fatal errors and out-of-memory conditions. In simplified terms, PHP could leave internal state pointing to memory that had already been freed. A later request handled by the same worker could reuse that stale state, creating opportunities for memory disclosure and memory corruption.
The two CVEs cover multiple underlying bugs:
CVE-2026-109618: An upload-progress unserializer use-after-free that can expose PHP worker memory and be developed into remote code execution.CVE-2026-109619: A fatal-error backtrace use-after-free and an OPcache type-confusion issue involving shared state, both of which can be developed into remote code execution.
Wiz combined the vulnerabilities into an exploit chain and demonstrated unauthenticated remote code execution against current WordPress, Drupal and Joomla containers. It also demonstrated that passwords, tokens and API keys could be read from PHP worker memory without first obtaining a shell.
The vulnerabilities are in PHP itself, not in WordPress, Drupal or Joomla. Those applications were simply used to demonstrate the exploits. Any application running on a vulnerable PHP version could be affected, even a basic Hello World app.
The age is what makes this particularly interesting. Wiz says the vulnerable code had been present for more than a decade. Its research also found that PHP versions going back to 5.4 were affected by at least one of the disclosed issues.
The discovery is interesting for another reason: AI played a substantial role in the research.
Wiz used its own Wiz Atlas platform alongside Gemini 4 Argon and GPT-5.6 Sol. It specifically credited Gemini 4 Argon with making important contributions to vulnerability discovery, exploit development, attack detection and patch verification.
Wiz says a single researcher completed in a week work that previously could have taken a specialist team months. The researchers worked with the PHP team, Calif and OpenAI on coordinated disclosure and patch verification. Cloudflare, AWS, Google, Fastly and other providers also prepared customer guidance and mitigations for deployments that could not immediately update.
The PHP project released:
- PHP 8.2.35
- PHP 8.3.36
- PHP 8.4.27
- PHP 8.5.12
There is an important distinction around exploitation status. On October 27, Wiz reported that none of the PHP2Shell exploitation attempts it had observed had completed successfully. By October 30, however, Microsoft was listing both CVEs as actively exploited and noting that CISA had added them to its Known Exploited Vulnerabilities catalog on October 26.
These are observations from different points in time and different sources. The earlier report is not a reason to treat the vulnerabilities as harmless.
The same week brought more runtime problems
Node.js
Node.js shipped security updates on October 21: 22.23.4, 24.22.1 and 26.12.1.
Among the fixes was CVE-2026-111497, an out-of-bounds write in Google’s Brotli decoder. An application that decompresses attacker-controlled Brotli data could suffer memory corruption, memory disclosure, a crash or potentially remote code execution.
This is not exclusively a Node.js problem. Brotli is used across a much wider software ecosystem, including .NET, PowerShell, PHP’s Windows builds, Visual Studio, Git for Windows and Python packages.
The affected Brotli library versions were 1.0.3 through 1.2.0. The upstream library’s fixed release is 1.2.1. Node’s October security releases included the fix, and later Node maintenance releases retained it.
GitLab
GitLab also shipped a critical patch on October 21 for CVE-2026-110264, a CVSS 10.0 vulnerability that could allow unauthenticated remote code execution on self-managed installations under certain conditions.
The flaw involved an integer overflow while compiling a specially crafted regular expression. It affected GitLab releases going back to version 7.0, first released in 2014.
The fixed releases were:
18.11.1319.0.1019.1.919.2.919.3.619.4.319.5.1
GitLab.com was already patched, but self-managed installations needed to upgrade.
GitLab’s official patch release.
One of eMudhra’s private keys was literally the number 4
This is the incident that makes you question how many layers of security can fail at once.
eMudhra had generated eight intermediate certificate authority keys during a ceremony on April 30, 2025.
These were P-384 elliptic-curve keys. A private key on that curve should be an enormous, randomly generated number. These weren’t.
On October 17, researcher Scott Helme reported the issue. On October 20, he published his technical findings, demonstrating that all eight private keys could be recovered from their public certificates on an ordinary laptop in under a second.
One key was 4. The other seven were 56, 66, 115, 155, 156, 230 and 237.
That’s eight bits of security where there should have been 384.
The affected certificates consisted of six TLS intermediates and two code-signing intermediates. Anyone recovering those private keys could potentially create certificates that appeared to chain back to a trusted authority.
eMudhra revoked all eight intermediate certificates on October 18. According to the company, seven had never been used in production, and the one that had been used had not issued a single certificate. Its infrastructure was not compromised, and it reported no data breach.
But that wasn’t the end of the impact. Browser root programs also distrusted the parent ECC root as a precaution. eMudhra said the resulting trust-store changes affected a US customer ecosystem with roughly 30,800 certificates under the broader G3 hierarchy, which it was transitioning to an alternative certificate path.
The explanation for how this happened is almost as concerning as the result.
eMudhra says a key wrap/unwrap step during the generation ceremony produced a second, weak set of keys alongside the intended strong keys. The weak keys were the ones that ended up in the browser root stores. The issue went undetected for around 18 months, including two independent audit cycles.
The company says its exact root cause is still under investigation. It appointed KPMG as an independent external reviewer on October 30.
The company’s incident statement is worth reading alongside Scott Helme’s technical breakdown.
eMudhra is also winding down its publicly trusted TLS certificate business. That does not mean it is abandoning its entire PKI, digital-signature or identity business. Its eSignature, digital-signature and enterprise solutions remain operational.
Tailscale
This wasn’t a breach of Tailscale either, but it did have consequences for Tailscale users.
Tailscale accepts TLS connections signed by certificates trusted by the system certificate store. The compromised eMudhra intermediates therefore mattered even though Tailscale itself was not involved in the original incident.
Its security bulletin explains that Tailscale version 1.104.2 or later includes Go 1.27.3, which explicitly distrusts the affected certificates even if they remain in the system certificate store.
In other words, the CA problem reached software that wasn’t even part of the original incident.
The Cost of a Stolen Key
Wiz traced two enterprise credentials exposed in one of the platform breaches to large-scale unauthorized use of Amazon Bedrock. In both cases, the unauthorized requests started within 30 minutes of the reported exposure.
One credential was used to make roughly 800,000 calls over 30 hours, with 90% succeeding. The other generated 1.7 million calls in 13 hours, with more than a million successful. Wiz estimates the combined usage exceeded $100,000 at list prices.
The traffic resembled coding-agent usage distributed through model resellers. Wiz identified two resale operations that had access to and tested the stolen keys, but couldn’t establish whether either was responsible for the original breaches or the high-volume traffic.
What to do if you are affected
If a platform or package you use is compromised, assume the credentials it could access are compromised too. Rotate them rather than trying to determine exactly which ones were stolen. This includes cloud credentials, database passwords, GitHub tokens, registry credentials and AI provider keys.
Once you’ve rotated them, check usage logs and billing for anything you don’t recognize. That can help you figure out whether the credentials were actually used before they were revoked.
For LiteLLM 1.107.0, don’t start Python in the affected environment; the malicious startup code can execute again. Rebuild from a clean version, follow LiteLLM’s cleanup guidance, and rotate the credentials that were accessible from that machine.
For the TensorLake and OVHcloud incidents, remove the affected releases and install clean versions. Rotate any credentials that may have been exposed. For the remaining vulnerabilities, apply the relevant vendor patches.
Conclusion
We’ve seen a lot of security incidents this year, and they are happening often enough that it is getting harder to treat each one as an isolated event. Shai-Hulud was one of the big ones. Black October is another. And there are still two months left in the year.
I think AI is becoming a pretty obvious part of this. It is making vulnerability research, exploitation and credential abuse faster and cheaper. We are going to see more of that, and I don’t think the next one or two years are going to be particularly quiet.
For a lot of these incidents, there isn’t much a developer can do once the thing they rely on gets compromised. A lot of the platforms and packages we use are getting hit too. At that point, the useful thing is knowing when something has happened, checking whether you’re affected, and patching or rotating keys as quickly as possible.
You can do everything right inside your own application and still end up exposed because a platform you depend on has a vulnerability that nobody knew about yesterday.
I don’t know what this looks like by the end of the decade. But I think it gets worse before it gets better.

