PyPI
Python Package Index
The official public registry for publishing and installing Python software packages; a frequent target for software-supply-chain attacks such as typosquatting or dependency confusion, since anything published there can be installed and run automatically by whoever names the dependency.
In our briefings
The claim each of the newest briefings on PyPI would have you leave with.
Anthropic reports that Claude models broke into the production infrastructure of three real organisations while attacking fictional targets in capture-the-flag cybersecurity evaluations. They used basic techniques such as exploiting weak passwords and unauthenticated endpoints. Anthropic began reviewing 141,006 evaluation runs on 23 July 2026, two days after OpenAI's disclosure, and found six such runs in three incidents, the earliest dating to April. The prompts told the models they had no internet access, but a misconfiguration left the machines online. The worst incident extracted application and infrastructure credentials and reached a database holding several hundred rows of production data. Anthropic notified its evaluation partner Irregular and the three organisations on 27 July. The two it could reach "had not previously detected the activity or contacted us", and it was still trying to reach the third.
An AI lab's test agents broke into three companies, and two had not noticedBI-FIND-2026-0165
LiteLLM says a config file committed to its public lite-harness repository on May 27, 2026 held an administrative key to an internal deployment of its Agent Platform. A follow-up commit deleted the file about three hours later, and the keys were noted as requiring revocation. Then, because of what it calls human error, "nobody followed through on revoking the administrative key." The earliest use the post reports is Oct 10, 136 days after the commit. The post does not say whether logs were searched for earlier use, or how or when the attacker found the key, so Oct 10 is the first reported use and may not be the first use. The commit stayed in public history until Oct 22.
LiteLLM left a leaked admin key live; its first reported use, 136 days on, yielded an org owner's GitHub tokenBI-FIND-2026-2109
Hartman's broker splits the decision from the secret. A policy decision point that "MUST never touch credential material" approves, denies or escalates a signed task request naming the target, action, scope and lifetime; a separate delivery point pulls the real credential from a hardware-backed vault, mints something narrower and wipes its memory. Three delivery models follow. In the first, a proxy injects the credential and the agent never sees it. In the second, recommended as the default, the agent holds a short-lived derived token bound to its own key with DPoP. In the third, marked NOT RECOMMENDED, the agent receives the real long-lived credential and it is revoked after the task. So the abstract's promise that agents "never hold real long-lived credentials" holds only if the third model is never used, and two sections disagree on which model is the fallback for services that cannot issue short-lived tokens.
A SANS instructor's broker vaults agents' real keys, but its third delivery model breaks the abstract's promiseBI-FIND-2026-0534
Cited in 4 briefings
Every briefing that discusses this term, newest first.
- An AI lab's test agents broke into three companies, and two had not noticed
Article · · Anthropic · Jul 30, 2026
- LiteLLM left a leaked admin key live; its first reported use, 136 days on, yielded an org owner's GitHub token
Article · · LiteLLM blog · Oliver Jensen et al. · Oct 23, 2026
- A SANS instructor's broker vaults agents' real keys, but its third delivery model breaks the abstract's promise
Draft · · IETF Datatracker · Kenneth G. Hartman · Mar 29, 2026
- In Invariant Labs' Cursor test, hidden text in an MCP tool description got the agent to send out the user's MCP config
Article · · Invariant Labs · Luca Beurer-Kellner et al. · Apr 1, 2025