
Anyone who claimed a renamed or deleted GitHub repository’s old name could publish to some packages; GitHub has since added a 48-hour expiry, but owners must check their settings.

A crafted pattern in a pipeline file could let an attacker without an account take over the server that holds a company's source code and build secrets.

Push Security says the lure page showed Anthropic's real install command while its copy button swapped in one that runs an attacker's script.

The DevOps company is buying selected Augment assets rather than the whole startup, aiming to hand AI-written code straight to its own delivery agents.

The 9.8-rated flaw sits in the cache server's multiprocess mode, and the project's own Kubernetes example exposes it on every network interface.

The new small model matches OpenAI's GPT-6 Luna on base price, adds effort levels from low to max and arrives with a 50% cut to Sonnet 5.5 cache reads.

The ModernBERT-based model judges a candidate secret in under 2 milliseconds and is in private preview, with general availability planned for later in October.

Announced at Team '26 Europe, the Agentic Multiplayer Protocol gives each agent an admin-assigned identity and scope, while Rovo Work takes on long, multi-step jobs.

The release also blocks shell characters in command-line usernames, hardens sftp downloads and promises faster releases as AI tools surface more bugs.

The Microsoft-owned platform is moving repository data into Azure Blob Storage and says internal tests show up to 35 times more write throughput.

GitHub rejected the report because a user must point the agent at untrusted content, but the researchers say the attack chain still works.

The companies used AI-assisted analysis to find the bugs and now let enterprises queue specific open-source dependencies for embargoed fixes and backports.

Checkmarx says one package, function-flag, has been delivering malware for more than a year and drew about 37,000 downloads with no advisory flagging it.

CVE-2026-21589 lets attackers without a login read files in the web root of self-hosted Jira, Confluence and Bitbucket, and Atlassian says exposed servers should come offline until patched.

The firm that made an early bet on Anthropic joined the AI coding startup's $5 billion round, a public show of support after last week's fight over a board adviser.

The count reflects both a surge in machine-found bugs and the kernel team's policy of tagging nearly every stable fix with a CVE, not 1,313 serious holes.

The company paused its open-source rewards program Oct. 1 after maintainers were flooded with AI-generated findings that did not hold up.

The Postgres company is betting that coding agents will need vast numbers of small, instantly created databases, and it is adding a SQLite-based engine to supply them.

A crafted flow configuration could break out of the prompt template sandbox and run commands on a server that holds JWT signing keys.

The Allen Institute for AI's open-source training framework targets trillion-parameter mixture-of-experts models and posts a throughput lead over Nvidia's Megatron-core in Ai2's own test.

Search results are billed at the partners' list prices from existing AI Gateway credits, and providers must follow Cloudflare's verified-bot and robots.txt rules.

The cheaper desktop AI system goes on sale Oct. 23 through six PC makers, and two linked units can run models of about 200 billion parameters.

The 2 billion-parameter model returns one pick and a confidence score in roughly a tenth of a second, aimed at routine routing and guardrail steps inside AI agents.

Some Google employees reportedly see a gap between the model's benchmark scores and its everyday work, a day after Gemini 4 Argon reached its first users.