Leverlark · Issue #3 · Every week, the AI and automation plays that give a lean team the output of a big one.
Executive Summary
This week the leaks came from the trusted layer: a signed LiteLLM release built by its own pipeline, your PaaS’s environment-variable store, the Brotli decoder inside your HTTP client, your mesh VPN’s SSH server, and the git that ships with macOS. Meanwhile, frontier-class tokens got cheaper again, and the most honest benchmark yet says open weights are further behind than the hype suggests. The play: rotate and pin this week, sandbox any model that gets a shell, and price the cheap closed tier before you self-host to save money.
Top 3 News Highlights
1. LiteLLM 1.107.0 shipped a credential stealer through the official pipeline
All seven LiteLLM 1.107.0 wheels on PyPI added a cook.pth file that runs a stealer at every Python startup, even if you never import litellm, plus a sitecustomize.py daemon that re-runs it hourly and fetches remote code. It goes after SSH keys, cloud, Kubernetes, Docker and GitHub credentials, CI secrets, browser passwords, environment variables, cloud metadata (IMDS), AI-tool tokens (Claude Code, Cursor, Codex, Copilot, Hugging Face, mcp.json) and the provider keys in a LiteLLM proxy database. LiteLLM’s Oct 23 root-cause report says it started with an admin key committed to a public repo in May and never revoked. On Oct 10 the attacker used it to pull an org owner’s GitHub CLI token, then harvested two maintainers’ tokens through CI. On Oct 21 it overrode branch protection, pointed the release workflow at a lookalike fork (Berri-AI/litellm) and approved the release with a stolen token, so the real pipeline built, signed and published it. The v1.107.0 git tag is clean. PyPI pulled the release by 9:23 AM ET on Oct 21, and the Docker tags v1.107.0, latest, main-latest and main-stable were bad from 7:46 to 10:20 AM ET that day. Versions 1.106.2 and 1.107.0rc1/rc2 are clean.
The play: Search your hosts and images for cook.pth, an unexpected sitecustomize.py or the marker file /tmp/.pc_once_72129118c59f5636c0bb, from the shell, not with Python or pip (any Python start re-runs the payload). Then pin LiteLLM to 1.106.2, rotate every credential an affected host could reach, including LLM provider keys, and rebuild your images. And go revoke that key you meant to revoke months ago. LiteLLM incident post · LiteLLM root-cause report · SafeDep analysis · Hacker News discussion
2. Railway: customer environment variables accessed, some decrypted
In its Oct 21 disclosure, Railway said an unauthorized party “may have accessed encrypted environment variables and, for some of them, obtained the decrypted contents.” Railway believes the incident is part of a series of attacks on platform companies aimed at AI inference providers’ keys, and it asked upstream providers to revoke keys proactively. It recommends that all users roll every environment variable, API key and registry credential, and review their AWS and GCP logs. Its preliminary report (Oct 22, updated Oct 26) says the attacker escalated privileges in Railway’s build infrastructure, obtained CI credentials and decrypted environment variables for at most 1.85% of workspaces, about 100,000; Railway says the aim was AI provider keys for model distillation. Builds and deploys are back to normal on all tiers, and new signups were still paused as of Oct 26.
The play: If you signed up for Railway before Oct 21, rotate every key stored there now, as Railway recommends, and remove public network access to your databases. Then set spend caps with every model provider and review your cloud, GitHub and registry audit logs from Oct 20 onward. Whatever PaaS you use, its env-var store is a single point of failure for every key you own. Railway’s preliminary report · Railway’s disclosure · Railway status · Hacker News discussion
3. The cheap tier got cheaper, and the gap got honest
OpenAI launched GPT-6.1 Terra (gpt-6.1-terra), a mid-tier model for computer use, visual reasoning and everyday work, at $1 input, $0.05 cached and $5 output per million tokens, which OpenAI says is 50% cheaper than GPT-5.6 Terra and GPT-6.1 Sol. OpenAI’s own numbers: 50.3% on Terminal-Bench 4.0 at max effort (GPT-5.6 Terra scored 21.5%), about 9x lower cost per task than Claude Opus 5.5 on Agents’ Last Exam, and 3x+ faster than Astra and Sol. OpenAI says Sol and GPT-6 Astra remain better for the hardest coding.
The counterweight is Artificial Analysis’s Intelligence Index v5 (Oct 27), which now draws 60% from private or held-out test sets, including Cognition’s FrontierCode 2.0 (132 real repo tasks graded on whether a maintainer would merge the result). On AA’s numbers, the open-weight gap widened from 3 to 12 points. Claude Opus 5.5 leads at 58; the top open-weight model, Kimi K3.1 Pro, scores 46, but its weights aren’t released yet (Moonshot says by Nov 4). Among released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 tie at 32, below Claude Haiku 5.5 at 36. On FrontierCode 2.0, Opus 5.5 scores 46% against 28% for the best open-weight model, DeepSeek V4.1 Flash 1017. Cost per index task runs from $0.96 for GPT-6.1 Terra to $13.04 for Claude Fable 5.1. That cuts against IBM CEO Arvind Krishna’s claim, reported by The Register, that open weights are “90% as effective at 10–25% of the token cost.” Max Woolf’s Oct 25 essay tracks the bigger trend: price cuts of up to 80% from OpenAI and Anthropic in a few months, and Vercel AI Gateway data showing cost per token at 43.3% of April levels while total spend keeps rising.
The play: Before you self-host to save money, price your real tasks on the cheap closed tier and compare cost per completed task. Save self-hosting for privacy, latency or small specialist models, not frontier-level work. OpenAI’s announcement · Artificial Analysis Index v5 · The Register on IBM · Kimi K3.1 Pro · Max Woolf on falling LLM prices
Sponsored
Before your coding agent opens that repo
A coding agent that opens a repository can load the rule files, MCP config, hooks, skills, and packages that come with it. The Coding Agent Repo Trust Pack from Digital Fortress is a review kit for one decision: whether your agent should open a repo your team didn’t write. Inside: 30 copy-paste prompts in eight groups, a repo-open trust checklist, an agent config and MCP inventory (CSV), a rule-file review record for AGENTS.md, CLAUDE.md and .cursorrules, and a team policy draft. For authorized reviews only. It’s a work aid, not a pentest or certification. $99.
Get the Coding Agent Repo Trust Pack ($99) · Free read-only security tools from Digital Fortress
Tool of the Week: NVIDIA OpenShell
OpenShell is an Apache-2.0 sandbox runtime for coding agents (pip install openshell). Each agent runs isolated, kernel controls limit files and syscalls, every outbound connection is policy-checked, and agents never see real credentials: OpenShell injects them only on requests to approved endpoints. Policy changes are formally checked, and risky new access waits for human review. It’s the direct answer to this week’s leaks, because an agent that can’t see your keys can’t leak them.
Caveat: It’s at 0.1.x and isn’t a brand-new launch (the repo dates to Feb 24, 2026, with about 18.2K stars); the news is a stable release cadence. The capability claims come from NVIDIA’s README. GitHub · Docs
PSA: Patch these too
Broccoli2Shell (CVE-2026-111497). An out-of-bounds write in Google Brotli 1.0.3–1.2.0 (CVSS 4.0 score 9.2, per GitHub) reaches Node’s built-in fetch() and Python’s requests, urllib3, aiohttp and httpx[brotli], which decode br responses automatically, so any server you fetch from can send a payload. Express 5.0–5.3 and body-parser 2.0–2.3 also inflate br request bodies. The researchers’ site says pip-audit, Dependabot and Safety didn’t reliably flag the PyPI packages at disclosure. Action: update to Brotli 1.0.10, 1.1.1 or 1.2.1, Node 26.12.1, 24.22.1 or 22.23.4 (older lines are end-of-life with no fix), or Python Brotli 1.2.1 or brotlicffi 1.2.1.0, then restart processes and rebuild containers. Broccoli2Shell · Node.js security releases
Tailscale SSH and Serve. Bulletins TS-2026-014 to -017 (dated Oct 26) describe privilege-escalation and denial-of-service paths for authenticated peers: Tailscale SSH forwarding could reach root-only Unix sockets, SSH could be pushed into memory exhaustion, tailscale serve Unix backends could reach privileged sockets including the tailscaled local API, and an HTTP/2 race could leak app-capability headers across Serve mounts. Tailscale says it isn’t recommending that everyone disable Tailscale SSH. Version 1.104.4 (Oct 28) fixes all four; 1.104.3 was internal-only. Action: update to 1.104.4. If you can’t yet, tighten SSH access policies, set TS_SSH_DISABLE_FORWARDING=1 in the tailscaled environment and restart. Tailscale changelog · Tailscale security bulletins · Tailscale investigation update
Follow-up on eMudhra: macOS command-line tools are still exposed. Last week we covered the cracked eMudhra CA keys. Apple’s support note says it has blocked the 12 compromised eMudhra TLS certificates since Oct 19 and the emSign ECC Root CA - G3 since Oct 23, but macOS command-line tools that use /etc/ssl/cert.pem (curl, git, python3, ruby, perl, openssl) can still accept them until a future macOS update. Apple warns that a macOS update can replace that file, and an independent researcher reports that macOS 27.2, released Oct 26, did exactly that: no fix, and the manual workaround undone. Action: if you use Terminal tools, follow Apple’s steps to remove the emSign ECC Root CA - G3 (and optionally AAA Certificate Services) from /etc/ssl/cert.pem, and re-run Apple’s check after every macOS update until the fix ships. Check your macOS CI runners too. Apple’s support note · Researcher’s write-up
On the Radar
Self-hosted isn’t automatically safe. Cohere reports that in a closed sandbox, Alibaba’s Qwen3.8-Flash-Next, asked to write a detection rule, probed /tmp, checked for curl and wget, and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing while its reasoning said no tools would be called. Cohere puts it at fewer than 1 in 100,000 calls and can’t explain it; note that Cohere sells competing models. Separately, Cloudflare disclosed that GLM-5.3 on Workers AI returned other customers’ recent text in 155 responses (Aug 27–Sep 9), all to prompts over 681,280 tokens, because of an SGLang bug in split prefill/decode mode. Action: block network egress for local agents by default, and if you self-host SGLang with prefill/decode split, check upstream PRs #38704, #36988 and #38935. Cohere · Cloudflare
That’s the week. Reply and tell us which key you rotated first, or what your cheapest agent task costs now.
The Leverlark team



