Update apt now: a Debian and Ubuntu bug lets a network attacker swap packages and run code as root
APT's maintainers fixed CVE-2026-109518, which lets a man-in-the-middle replace a package after it was verified and run code as root on Debian and Ubuntu. OpenAI is buying Pierre, the git host behind Lovable and Bolt, for Codex; Google previews AlloyDB Express, a scale-to-zero Postgres for coding agents; and Render opens sandboxes for agents.
- APT, the package manager on Debian and Ubuntu, had a critical bug (CVE-2026-109518): a man-in-the-middle could replace a package after verification and run code as root. Upstream fixes cover every stable APT branch, and Debian and Ubuntu updates are rolling out.
- OpenAI is buying Pierre Computer Company, whose Code.Storage stores agent-generated code for Lovable, Bolt and Amp, to expand Codex. OpenAI’s announcement summary gives no price.
- Google’s AlloyDB Express, in preview, creates a PostgreSQL 18 database in about two seconds and pauses it after 15 idle minutes, at $0.115 per compute-unit hour.
- Render Sandboxes are in public beta: Firecracker microVMs for agents that bill CPU only while it’s busy ($0.12 per active CPU-hour) and suspend when idle.
- The Kimi K3.1 Pro weights are out on Hugging Face (2.8 trillion parameters, 104 billion active), as Moonshot promised at launch.
Update apt before your next install
APT’s maintainers on Wednesday shipped fixes for CVE-2026-109518, a critical bug that lets an attacker on the network replace a package after APT has verified it. The attack combines redirects with transient server errors that make APT retry a download, so it overwrites a verified file with the attacker’s payload. Packages install as root, so that means code execution as root. Max Justicz found it. It needs Acquire::Retries, which has been on by default since APT 2.3.2; older systems are exposed too if scripts turned it on. Debian’s DSA-6607-1 fixes trixie (APT 3.0.4); bookworm gets 2.6.2 through Debian LTS (DLA-4880-1), and Debian’s tracker still listed forky as vulnerable when we checked. Fixes cover every stable branch from 1.2 to 3.2, and the post says Ubuntu updates are being released.
Why it matters
Every Debian or Ubuntu server, CI runner and Docker image that runs apt-get install is exposed whenever it downloads over a network someone else controls. Upgrade the apt package first, rebuild base images, and pin CI images to patched tags.
Package signing didn’t save anyone
APT trusts a package because its hash comes from a signed index, but this bug swaps the file after the check. Justicz also found apt’s 2019 redirect bug, CVE-2019-3462.
OpenAI buys the git host behind Lovable and Bolt
OpenAI will acquire Pierre Computer Company, “expanding Codex’s tools for building, understanding, and reviewing code”. Pierre makes Code.Storage, git hosting for machine-generated code. Lovable, Poke, Bolt, Amp and others use it to store generated code, agent memory files and other artifacts, with repos up to 32TB and no cap on repo count. OpenAI’s announcement summary gives no price and doesn’t say what happens to Code.Storage customers.
Why it matters
If your product keeps user code or agent state in Code.Storage, your storage vendor is about to belong to OpenAI. Ask Pierre for its plans, and check how easily you can export to plain git.
OpenAI is buying plumbing its rivals use
Amp is a coding agent like Codex, and Lovable and Bolt sell app building from prompts, a market OpenAI could move into. OpenAI gets git storage built for agent workloads at large scale. If OpenAI doesn’t commit to serving outside customers, expect some to move off it.
Google’s Postgres for agents pauses itself
Google Cloud launched AlloyDB Express in preview, a PostgreSQL service built for developers and coding agents. A database is ready in about two seconds (up to 20 with Private Service Connect) through the gcloud CLI, REST API, MCP or console. It pauses after 15 idle minutes and resumes in about 20 seconds on the next query. Compute costs $0.115 per compute unit (CU) hour, where one CU is one vCPU with 4 GB of RAM, and storage $0.30 per GB-month. The docs cap the preview at 2 vCPUs and 20 GB of storage. It runs in seven regions, with no backups, point-in-time recovery or SLA. The free tier gives 50 compute units and 2 GB of storage a month for the first 12 months.
Why it matters
Agents can now give every test run or prototype its own isolated Postgres inside a Google Cloud org, with IAM and network policies applied automatically. A one-click upgrade moves a database to full AlloyDB.
Neon’s price, Google’s guardrails
| AlloyDB Express | Neon (Launch) | |
|---|---|---|
| Compute | $0.115 per CU-hour | $0.106 per CU-hour |
| Storage | $0.30 per GB-month | $0.35 per GB-month |
| Smallest size | 1 CU | 0.25 CU |
| Pauses after, wakes in | 15 minutes, about 20 seconds | 5 minutes, a few hundred ms |
| Backups | None in preview | Restore history ($0.20 per GB-month) |
Neon’s CU is about 4 GB of RAM, so the units are close. AlloyDB Express is cheaper on storage, not compute, and Neon scales smaller and wakes faster. Google’s case is placement inside Google Cloud’s security controls. Our database-per-agent comparison has the wider field.
Render’s sandboxes bill only busy CPU
Render launched Render Sandboxes in public beta: isolated Firecracker microVMs where agents install packages, edit files and run code. Each sandbox can run indefinitely, snapshots and suspends after 15 idle minutes, and is typically ready in under a second. Sandboxes can allow all outbound traffic, a domain list, or none. Credential brokering, which adds API keys outside the sandbox so code inside never sees them, is in early access. Pricing is $0.12 per active CPU-hour plus $0.038 an hour per 2 GB of memory, and snapshots cost $0.08 per GB-month. It’s available in Oregon only.
Why it matters
Teams already on Render can run agent workloads in the same workspace as their web services and databases, under one bill and one API key system. Private networking to those services isn’t in the beta yet.
Mid-priced, unless the CPU stays busy
At the basis of our sandbox comparison (2 vCPU, 4 GB, CPU a quarter busy), Render costs about $0.136 an hour. That’s below E2B’s $0.166 and Vercel’s $0.149, and above Fly.io Sprites’ $0.107. Fully busy, it rises to about $0.316, nearly twice E2B but below Vercel’s $0.341. One region and beta limits make it a convenience for Render customers, not a category winner yet.
Quick hits
- The Kimi K3.1 Pro weights are on Hugging Face under Moonshot’s own Kimi K3.1 Pro License: 2.8 trillion parameters, 104 billion active, 1M context, MXFP4 weights. Vercel’s AI Gateway added more zero-data-retention providers; until now Moonshot was the only provider there. See Kimi K3.1 Pro’s price and benchmarks.
- VS Code 1.145 lets nested agent sessions use different harnesses in preview, so Claude can review a plan Copilot wrote in the same session. It also adds cron schedules for local automations.
- Cloudflare added per-agent controls for personal AI agents: a Personal Assistants bot category, per-agent traffic analytics, and policies to allow an agent, block it on ad pages, or block it.
- Lambda plans a $33.1 billion data center campus in Lowndes County, Alabama, with construction from early 2027; it gave no capacity figure.
- Data center limits won at the midterms: The Verge, citing Redbud Consulting’s tally, says 33 of 45 local ballot measures will restrict development. Texas governor-elect Gina Hinojosa backs a statewide moratorium.