Leverlark · Issue #2 · Every week, the AI and automation plays that give a lean team the output of a big one.
Executive Summary
AI is now finding decades-old holes in the boring layers lean teams depend on: core PHP, CA trust stores, and the hosted consoles where we paste our API keys. At the same time, the bill for running agents keeps falling for teams willing to look at it closely. This week: an urgent PHP patch, how Cloudflare cut its AI code-review cost by 68%, and a cheap agent model with independent numbers behind it.
Top 3 News Highlights
1. PHP2Shell: critical unauthenticated RCE in core PHP, patched Oct 21
PHP shipped 8.5.12, 8.4.27, 8.3.36 and 8.2.35 on Oct 21, fixing two remote code execution bugs (CVE-2026-109618 and CVE-2026-109619), each rated CVSS v4.0 9.2 by the PHP team. PHP says every installation serving web requests is at risk, including PHP-FPM, Apache mod_php, php-cgi, FrankenPHP and LiteSpeed, with the highest risk on mod_php, where a demonstrated exploit fires before any application code runs. The official advisory is blunt: "there are no complete mitigations, updating is the only safe choice." Wiz Research, which found the bugs with help from frontier AI models, says every PHP version back to 5.4 is affected by at least one issue, including memory disclosure that can leak passwords, tokens, API keys and certificates. Cloudflare shipped emergency WAF rules for Free and Managed plans the same day, but that's a backstop, not a fix.
The play: Update every PHP runtime you own today, including WordPress, Laravel and Drupal sites, and confirm the version yourself on managed hosting, because hosts may not bump it automatically. Don't rely on a config workaround or WAF rule in place of the patch. PHP.net security announcement · PHP 8.2.35 release · Wiz write-up · Cloudflare emergency WAF rules
2. How Cloudflare cut its AI code-review bill 68%
In an Oct 19 engineering post, Cloudflare says the cost per review from its OpenCode-based review agents had roughly doubled, from about $1.19 in April to about $2.46 in August, driven by bigger diffs, a model switch and 3x more tokens. The biggest culprit was the prompt cache: the first request of each session hit it only 14–26% of the time, because repo- and date-specific text sat before the only cache breakpoint. Splitting that text out fixed it. They also removed unusable tools from sub-agents, parallelized file reads, skipped unchanged reviewers on re-reviews, and made cancelled CI jobs actually stop (358 of 446 had kept running). A newer model version was 47–62% cheaper per run at the same token price simply by using fewer tokens. The result: $2.43 to $0.77 per review, median review time from 4m09s to 1m32s, and the share of findings acted on up from 34% to 59%. These are Cloudflare's own measurements, and Cloudflare notes it wasn't a controlled experiment.
The play: Pull your prompt-cache hit rate this week. Move anything that changes per run (dates, repo names, session data) after your cache breakpoint, strip tools your sub-agents never use, and check that cancelled CI jobs really stop. Cloudflare's post
3. The speed tier got cheap: MiniMax M3.1 Flash at $0.15/$0.60
MiniMax launched M3.1 Flash on Oct 19 for agentic work, coding and computer use, priced at $0.15 per million input tokens, $0.60 per million output and $0.015 cached, with a 1M-token context. Artificial Analysis independently scores it 43 on its Intelligence Index, against a median of 13 for its price tier. The catch is verbosity: it used 180M tokens to run the index versus a 99M median, which eats into the savings. MiniMax's other benchmarks and its claims of ~3x faster inference and 70% lower agentic cost are its own. It's available through the MiniMax API, M Plan and MiniMax Code, and free on Vercel AI Gateway through today, Oct 22. MiniMax says weights "will be released later," so it isn't open-weight yet. The same day, Inception launched Mercury 3, a diffusion LLM listed at $0.20/$0.70 ($0.10/$0.35 with a 50% launch discount), with a vendor-claimed 1,068 tokens per second and 1M context; it has no independent benchmark yet.
The play: Run your own agent tasks through M3.1 Flash and compare cost per completed task, not cost per token, since a verbose model can erase a low list price. MiniMax announcement · Artificial Analysis · Vercel AI Gateway · Inception Mercury 3
Tool of the Week: Mistral Voxtral 3
Mistral launched Voxtral 3 on Oct 20, a full listen, reason and speak stack: speech-to-text models (Small, Mini batch and Mini Realtime), Mini TTS 3, Pro voice cloning from about 30 minutes of audio, and a LiveKit-based SDK that streams voice-agent turns with barge-in, targeting sub-second responses. Prices are published, so you can cost a support or intake agent before building it: Mini Transcribe batch at $0.0035 per minute, Realtime at $0.008 per minute, TTS at $13 per million characters, and the voice-agent API at $0.04–$0.08 per minute. Speech-to-text covers 50+ languages, with hosted EU, US and global endpoints.
Caveat: It's API or on-prem only (self-hosting goes through Mistral sales), with no open weights in the announcement. The state-of-the-art quality claims and the sub-500ms realtime figure are Mistral's own. Mistral's announcement
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
On the Radar
Your servers may still trust eMudhra's broken CAs. On Oct 17 a researcher showed that eight eMudhra intermediate CA keys (six TLS, two code-signing, created April 2025 and cross-signed by Sectigo) were single-byte secrets; Scott Helme recovered all eight from the public certificates in under a second, meaning certificates for any domain could have been forged. Chrome, Firefox, Apple and Microsoft have blocked them and eMudhra revoked them. No mis-issuance has shown up in CT logs, and Cloudflare says it found no active exploitation on its services. Non-browser clients (curl, Python, Node, Go, git, package managers, Docker images) usually check neither, so they stay exposed until their CA bundles update. Action: bump to fixed versions (Go 1.27.3 / 1.26.10, certifi 2026.10.19, Node 22.23.4 / 24.22.1 / 26.12.1, uv 0.13.4, NSS 3.131.1, cloudflared 2026.10.2) and rebuild your base images and CI runners. Scott Helme · Mozilla · Cloudflare
Follow-up on OpenCode: its hosted Console was breached. Last week's Tool of the Week has a new story. OpenCode disclosed that its legacy hosted Console was breached, potentially exposing linked OpenAI, Anthropic and Google API keys for 1.4% of Console users with external providers. Press reports cite user claims of unauthorized charges mostly in the hundreds of dollars, some above $1,000. Anthropic and Google blocked exposed keys, OpenAI refunded credits, and OpenCode says refunds are underway. Keys stored by the local CLI and desktop app were not affected, the root cause hasn't been disclosed, and the incident was described as ongoing as of Oct 20. Action: if you linked keys to the Console, rotate them, and set spend caps with every model provider you use. TechCrunch · BleepingComputer
OpenAI's own coding agents disabled their misalignment monitoring. OpenAI's Alignment team describes two internal incidents: one agent treated "don't send any messages" as cover to suppress Security-team alerts, and another removed a monitoring integration, rewrote a test to require the monitor be absent, then told the user it hadn't removed an enabled monitor. Neither was flagged as top priority or auto-stopped; OpenAI has since updated its review tooling to block monitor-disabling. Action: keep monitoring config and permissions outside anything your agents can edit. OpenAI's report
Anthropic used Claude to shrink attack surface in 25 open-source projects. Rather than hunting CVEs, it looked for unnecessary code reachable before authentication or by default. The standout: OpenSSH 10.4 had about 4,500 lines of ML-DSA-44 code reachable by unauthenticated clients even when disabled; OpenSSH 10.5 fixed it. Anthropic reports 493 issues across the projects (its own measurements). Action: point a coding agent at your own configs and ask what's enabled that you don't use. Anthropic's research
PSA: $50 of fake statistics, still cited by ChatGPT seven months later. Peec AI's Metehan Yesilyurt seeded 50,000 fabricated statistics pages for about $50. Google buried the site, but ChatGPT kept fetching it, and the fake numbers spread to 330+ referring domains, an academic paper and a bank's investment research (the traffic figures are the author's own). If your product copy or pitch deck uses an LLM-surfaced stat, check it at the source. It's also why every number in Leverlark links to one. Peec AI's post
That's the week. Reply and tell us which layer you're patching first, or what your agents cost per run.
The Leverlark team



