Newsletter

Credentials of Thousands of Organizations Stolen in LiteLLM Supply-Chain Attack

A poisoned open-source scanner rode into the LiteLLM AI gateway and handed an attacker cloud, SSH, and Kubernetes keys belonging to more than 2,500 organizations, while a Pew study finds over a third of post-ChatGPT web pages show signs of AI authorship.

A wide view down an aisle of server racks in a computer-room, racks receding into blue-tinged distance with rows of colored status lights.
Lead image Server racks in the NOIRLab headquarters computer room in Tucson, Arizona, illustrative of the enterprise infrastructure whose cloud and server credentials were targeted in the LiteLLM supply-chain attack. Image: NOIRLab/NSF/AURA/T. Slovinský, CC BY 4.0. · Image: NOIRLab/NSF/AURA/T. Slovinský · CC BY 4.0
On this page

Security vendors CloudSEK and Hudson Rock report that a supply-chain attack on LiteLLM — an open-source gateway that translates API calls for more than 100 large language models into a single OpenAI-compatible format — harvested credentials from more than 2,500 organizations, among them Cisco, Samsung, AWS, Airbus U.S. Space & Defense, Thales, and the London Stock Exchange Group. The operation began on March 24, 2026, when a financially motivated group the firms call TeamPCP compromised Trivy, Aqua Security’s open-source vulnerability scanner; LiteLLM then brought the poisoned build of the trusted tool into its own supply chain without verifying its identity, granting the attacker server administrator privileges and installing a credential-stealing implant. The list of victims includes some of the world’s most sensitive organizations, and the stolen material reportedly reached beyond corporate data into cloud keys, SSH keys, Kubernetes tokens, environment variables, repository and package-publishing tokens, and AI-provider keys.

The report, covered by TechRadar, stresses that such secrets act as keys to many systems rather than a single compromise, and that some credentials were still reported working months after the original breach, leaving a lingering risk until they are changed. These figures come from the two vendors’ investigations and have not been independently confirmed.

Featured source: TechRadar .

GitHub Details the August 17 Outage and the Reliability Work Ahead

GitHub’s CTO published a post-mortem of the August 17 outage that lasted 7 hours and 47 minutes, disrupting github.com, authentication, Actions, APIs, pull requests, issues, and Copilot. The root cause was a capacity failure: a critical infrastructure component in the company’s Central US data center failed to scale with a new traffic peak, producing cascading authentication failures that were made worse by a client-side retry loop in Copilot services. GitHub says it has since added over 3 million CPU cores and 120 petabytes of storage, and that Azure now serves roughly 58 percent of platform load, up from 12 percent in May, as it applies consistent retry limits to stop retry storms.

Filed from: GitHub Blog .

Pew Finds Over a Third of Post-ChatGPT Web Pages Show AI Authorship Signs

A Pew Research study built on Common Crawl and Open Pangram’s detection technology found that more than one-third of web pages published after ChatGPT’s November 2022 launch show signs of having been written or substantially edited by AI. In a random sample of 10,000 pages collected in July 2026, about 10 percent showed “significant signs of AI authorship,” a figure that rose to 35 percent once older, pre-ChatGPT pages were removed from the sample. Pages on .com domains showed signs of AI authorship at roughly 10 times the rate of .edu or .gov pages.

Filed from: TechCrunch , Pew Research Center .

Cisco Patches Four Critical Flaws in Secure Workload Software

Cisco disclosed that Secure Workload Software, its micro-segmentation product formerly known as Tetration, contains four critical vulnerabilities and one high-severity bug. The two highest-rated issues include improper access control and missing authentication or authentication bypass (CVE-2026-20315, CVE-2026-20317), a 9.9-rated flaw involves improper neutralization covering command, OS, and argument injection (CVE-2026-20231), and a 9.6-rated issue is an improper input-validation problem (CVE-2026-20318). Cisco said even users of the software’s SaaS offering have updates to install.

Filed from: The Register .

OpenAI Unveils Private Safety Processing for Zero-Data-Retention Customers

OpenAI announced Private Safety Processing, a mechanism to automatically scan customer model interactions for safety risks without breaking Zero Data Retention (ZDR) commitments, in a move The Register frames as chasing Anthropic’s enterprise customers. Under the approach, ZDR deployments can run on customer-controlled infrastructure or on OpenAI infrastructure protected by customer-controlled encryption keys, and automated systems can identify potential misuse and return limited safety signals without exposing underlying prompts or responses to OpenAI personnel. Anthropic, by contrast, retains inputs and outputs for its covered top-tier models for 30 days to support safety work even under ZDR; OpenAI has not yet published full technical details of its mechanism.

Filed from: The Register .

Apollo Confirms Data Breach Amid Hacking Wave on Private Equity

Private equity firm Apollo Global Management confirmed a data breach in which hackers used a social engineering attack to reach its cloud environment between July 6 and July 10, stealing names, birth dates, contact information, and Social Security numbers. The company — the world’s largest private equity firm with $938 billion in assets under management and roughly 5,000 employees — disclosed the incident in a letter filed with California’s attorney general, without specifying whose data was taken or whether a ransom was paid. The confirmed breach follows warnings from Google researchers that hackers were targeting private equity and financial giants in a widespread extortion campaign.

Filed from: TechCrunch .

Roblox Agrees to Court-Enforceable Child-Safety Changes After Regulator Finds Gaps

Australia’s eSafety Commissioner Julie Inman Grant said Roblox is still failing to stop adults from creeping on kids, after agency testing found adult strangers could send connection requests to children, that children’s profiles and contact lists were visible to anyone with no option to restrict them, and that such contact did not trigger parental alerts. Roblox, which about 1.7 million Australian children use, has entered a court-enforceable agreement to make changes within three months — including making children’s accounts private by default, preventing adults from contacting minors without parental consent, and reporting outcomes after complaints — and is the first platform to submit to independent audits under Australia’s Online Safety Act.

Filed from: Ars Technica .

From the Community

Micron Unveils $10 Billion Research Hub in Boise

Micron Technology announced Micron Research Labs, a U.S.-based research institution headquartered in Boise, backed by a planned $10 billion investment over the next decade and focused on memory technologies, advanced architectures, packaging, and future semiconductor manufacturing. The company plans to break ground in 2027 on a facility designed to host hundreds of researchers, and describes the hub as building on its previously announced $250 billion U.S. manufacturing and R&D commitment.

Filed from: Micron press release .

DeepSeek Releases a Vision-Capable Model

DeepSeek introduced an experimental vision model, deepseek-v4-flash-vision-exp, that accepts images alongside text through its API, supporting JPEG, PNG, GIF, and WebP formats via base64 inline, external URLs, or the Files API. Images are resized to a maximum of roughly 800×800 pixels, capping token cost at 384 tokens per image, and the model is available through OpenAI-compatible, Anthropic-compatible, and Responses API endpoints, though it supports images only in user messages.

Filed from: DeepSeek API documentation .

Huzzah: a Declarative Pseudocode Editor for AI Coding

Developer Daniel Vaughn introduced Huzzah, an experimental editor that replaces long-form imperative prompts with persistent, declarative pseudocode files. When a user saves a .hz file, Huzzah captures the diff and uses it as the prompt to an LLM, regenerating the affected source code. The approach aims to preserve human intent, reduce token consumption, and make prompts serve as documentation, though the author notes limitations including suitability mainly for new codebases and difficulty expressing cross-file dependencies.

Filed from: Daniel Vaughn’s blog .

Vomit: a Tool to Clean Up Claude’s Verbose Token Output

A new open-source tool called Vomit pipes a model’s verbose token output through a local LLM to translate it into concise English. It is fully local, with no telemetry and no external dependencies, and offers a non-invasive mode to run alongside an existing session. The author notes the local LLM can only see what the model tries to communicate, so it may hallucinate and risk missing the original message, and describes the tool as “vibe-coded” and only tested on its intended platform.

Filed from: Vomit repository .

Continue reading

Complete index →