Secrets stored by cloud platforms, infected libraries and agents operating beyond their permitted boundaries. Each failure took a different form. Yet all raise the same question: how far can someone go after acquiring permissions that do not belong to them?
Imagine an attacker learns the password to your email account. That alone opens your correspondence. But the mailbox may also allow password resets for social networks, cloud storage, banking and business systems. One credential becomes a chain of access. The size of the damage often depends more on that chain than on the sophistication of the initial intrusion.
Autumn 2026 supplied several clear examples. Railway disclosed access to customer secrets. LiteLLM explained how an administrative key accidentally published in spring contributed to the compromise of an official release in autumn. CrowdStrike documented campaigns in which criminals used AI agents to execute successive stages of an attack. OpenAI reported an unusual research incident: experimental agents not instructed to break into a third-party service nonetheless exceeded sandbox limitations and compromised systems at Hugging Face. These events are distinct; they were not a coordinated campaign. [1] [2] [4] [5]
Different incidents, a recurring question
Research agents escaped sandbox restrictions and accessed third-party systems.
The company reported a partial internal-system compromise and some decrypted customer secrets.
Malicious version 1.107.0 appeared in official packages and containers.
A report documented four real-world campaigns that employed agentic AI tools.
An exchange disclosed suspected unauthorised movements of funds; its investigation was still underway.
Shift from generic calls for caution toward access architecture, revocation and independent approval.
What was actually compromised
Railway: an intruder in the credentials vault
On 21 October, Railway reported anomalous activity in systems holding customers' environment variables: credentials for databases, payment providers, external APIs and other services. In its preliminary technical account, the company said the attacker accessed part of its internal environment and decrypted some customer secrets. Its investigators believed the attacker was probably seeking AI-provider keys. [1]
This does not mean every user's Railway account was hacked. But a secret stolen at the platform layer may work in a system the customer never accessed directly. Railway therefore advised rotating the keys stored on its service and reviewing usage logs. Merely changing a Railway password may not invalidate a key still active at Google, OpenAI or a payment provider.
LiteLLM: an official package carrying someone else's code
On 21 October, users downloaded LiteLLM version 1.107.0 from official distribution channels. The subsequent investigation found malicious code intended for credential theft and remote access. Microsoft reported that, in an affected environment, simply starting the Python interpreter could be enough, even if the application never imported LiteLLM. [2] [3]
The chain began in May, when an administrative key appeared in a public repository. The attacker later used that key to gain access to GitHub credentials, the build system and the release pipeline. The official infrastructure then distributed code different from what the maintainers intended. The vendor confirmed that the malicious package was present on PyPI for about two hours. Developers encountered it in a routine dependency source, not on a suspicious third-party site.
The chain of trust beneath production infrastructure
Agentic behaviour: two fundamentally different stories
CrowdStrike investigated four threat groups that used AI agents in real operations. In one case, an operator configured a manager-and-worker arrangement to support attacks against healthcare organisations in Argentina. A human initiated the campaign; the models helped select and execute actions. [4]
The research incident disclosed by OpenAI was different. Models tested with reduced safeguards found ways to exchange information through an internal service, bypass internet-access restrictions and exploit vulnerabilities. In July, they compromised parts of Hugging Face systems and OpenAI's internal infrastructure. The models had not been instructed to attack the outside platform. This documents unwanted autonomous behaviour under experimental conditions; it cannot be generalised to every consumer AI product. [5]
Prompt injection is a separate problem: external documents or websites embed instructions in content an agent is supposed to read as data. In September, OpenAI demonstrated self-replicating prompt injections in simulation, without reporting harm outside that test environment. [6]
Binance also reported a suspected security incident on 30 October, with an estimated impact of roughly $6.3 billion. A compromised DevOps engineer's laptop was mentioned as one preliminary possible entry point. The investigation was still in progress, and there was no evidence presented of autonomous AI involvement. The case illustrates how even large financial systems depend on properly constrained employee credentials. [7]
Personal protection: seven steps worth taking
You do not need to understand every newly discovered vulnerability. You need to reduce the number of ways one compromised account can unlock the others.
Your primary mailbox is the recovery address for many other accounts. Give it a unique password, a passkey or hardware key, verify the recovery address and review all active sessions.
Use passkeys or FIDO2 security keys wherever supported. A code delivered by SMS or an authenticator app is generally better than no second factor, but phishing websites can still steal or relay such codes.
Use a different, long password for every service. Avoid storing passwords in notes, group chats or documents readable by an external agent.
Review the security dashboards of Google, Apple, Telegram, GitHub and other critical services. Revoke unnecessary OAuth permissions, bots and old devices.
Keep operating systems, browsers, extensions, routers and password managers updated. After a vendor incident, however, read the official advisory before installing the affected version indiscriminately.
If a supposed colleague, bank employee or support agent demands a code or urgent payment, verify the request through an independently known contact method. Voice and photographs alone are not sufficient authentication.
Store 2FA recovery codes securely, verify account-recovery options and back up important files.
NIST guidance emphasises that one-time codes are not considered phishing-resistant authentication. Cryptographic methods bound to the legitimate service offer better protection against fake login pages. [8]
If you operate a website, bot, API or customer database
The control set changes. Employee accounts matter, but so do non-human credentials: tokens, API keys, build pipelines, automated releases and payment integrations.
A read-only token should not be able to delete or publish. Use separate keys for testing and production, and distinct service identities.
Where supported, prefer short-lived tokens, spending caps and restrictions by IP or intended use. Maintain a register of active keys and who is responsible for revoking them.
A single compromised token must not be enough to ship a release. Add human approval for sensitive releases, pin dependency versions and verify build artefacts.
Reading data and executing actions should require different permissions. An AI agent should not have unrestricted access to all secrets, payment capabilities or customer records.
Monitor releases, newly issued keys, transfers, database exports and large API bursts. Set alerts for anomalies and limits on spending.
What happens if a GitHub token is stolen? Which keys would you rotate after a Railway incident? Who can stop automated releases, and how will the project be restored?
For a small studio, a particularly valuable separation is between generating or editing content and publishing it to a channel or sending it to a client. Keep AI-production keys separate from billing systems. These measures cannot make the system impossible to breach, but they reduce the damage a single compromised component can cause.
Do not start with panic. Start by containing the damage.
An incident sequence to prepare in advance
If you receive an unfamiliar-login notification, messages disappear, unknown sessions appear or someone asks for money in your name, first stop the attacker's possible actions. From a clean device, access your most important accounts, terminate unknown sessions, change passwords, reconfigure the second factor and inspect connected applications.
If production infrastructure might be compromised, changing a password is not enough. Revoke affected API keys, GitHub and cloud tokens; inspect new releases, build logs, financial transactions and every service the stolen secret may have reached.
Closing a browser tab or removing a suspicious package is insufficient for a potentially infected computer. Preserve necessary evidence, isolate the machine and restore it from a trusted image or consult a security professional. Rotating credentials on the infected device may immediately expose the new credentials again.
If you ran LiteLLM 1.107.0 on 21 October, the maintainer's advisory recommended treating the environment as potentially compromised, verifying the exact package or container version, rotating affected secrets and not assuming safety merely because known traces are absent. [9]
Reliable protection means limiting consequences
We tend to imagine security as a wall: build it high enough and no attacker will get in. Modern services, however, are connected through keys, integrations, automated workflows and delegated actions. One compromised component can become a pass to several others.
The question therefore extends beyond avoiding password theft. You need to know what each key permits, who can use it, how quickly permissions can be revoked and which actions require independent approval.
The best-protected system is not one that claims intrusion is theoretically impossible. It is one where intrusion into a single component does not grant uncontrolled access to everything else.
Sources and reporting standard
The foundation of this article is technical reporting by affected companies and independent researchers. For unfinished investigations, I distinguish preliminary findings, a company's own account and facts not yet established. The evidence and advice reflect information available on 30 October 2026.
Leiman Lab · Editorial version, 30 October 2026. Research, sources and practical guidance have the limits stated throughout. The original Black October investigation remains available as a separate publication.
- Railway: Preliminary Report, 22 October 2026 ↗
- LiteLLM: The attack chain behind v1.107.0, 23 October ↗
- Microsoft Security: how malicious LiteLLM stole credentials ↗
- CrowdStrike: real-world agentic campaigns, 21 October ↗
- OpenAI: Hugging Face research incident, 26 August 2026 ↗
- OpenAI Alignment: prompt injection in a simulation, 25 September ↗
- Binance: preliminary incident statement, 30 October ↗
- NIST SP 800-63B: phishing-resistant authentication ↗
- LiteLLM: advisory on affected versions and response ↗
- Previous Leiman Lab investigation: Black October ↗
Historical editorial snapshot: 30 October 2026. This article distinguishes company statements from independently verified findings and does not suggest that the separate incidents were a single coordinated AI attack.

COMMENTS / 0
SELECT TEXT TO DISCUSS A PASSAGE