In partnership with
TOGETHER WITH TODAY'S PARTNER
Your AI budget tripled. See real usage patterns with Harmonic.
AI spend just became one of the biggest line items on the P&L. And most teams can't say what it's actually producing.
Harmonic Security maps AI activity to the use cases and teams driving it, showing which tools are earning real productivity, which are shelfware, where sensitive data is going, and how adoption is trending over time. Across approved AND unapproved AI tools.
That's the data you need to show your board exactly where the return is coming from.
Good morning, {{first_name|there}}. An AI security tool fixed one bug in Snowflake's repo and quietly opened a worse one — and an AI attacker found it five days later.
Read time: 3 minutes. Same time, every weekday — rate today's issue at the bottom.
🚀 The Big Story: Copilot's security fix opened a worse hole
Wiz disclosed that a GitHub Copilot Autofix patch merged into a Snowflake repo closed one vulnerability and introduced a command-injection flaw in the same pull request — one an autonomous agent exploited within five days.
The patch traded safe for unsafe: PR #1218 replaced a hardened env-variable pattern with a GitHub issue title interpolated straight into a shell command. A single apostrophe becomes code execution.
Five days, not five months: merged June 18, found and exploited June 23 by Wiz's Red Agent using a malicious issue title. Snowflake patched and rotated credentials the same day.
What leaked: a Jira API token, service email, and base URL — enough for read access across engineering, security compliance, and bug bounty projects.
The input was public: anyone can open an issue on a public repo. The attack surface was a text field, and the fix that created it carried a security tool's badge.
Jason's take: The failure isn't that a model wrote bad code — models write bad code constantly and we catch it. The failure is that this patch arrived wearing a security tool's uniform, so it got less review than a junior dev's PR would have. Meanwhile the discovery clock collapsed from months to five days, because the attacker automated too. If you're merging AI-suggested patches, you now need an AI adversary pointed at your own repos — the speed asymmetry runs one direction, and it isn't yours.
⚡ Quick Hits
Groq raised $350M at $3.5B — half its old price. Down from $6.9B in September 2025, with Nvidia joining the round after licensing its tech and hiring founder Jonathan Ross. The chip startup is now an inference cloud.
Google shipped Gemini 3.7 Flash three weeks after 3.6. FrontierCode 1.1 jumped 34.4% to 43.6% and DeepSWE 49% to 65.3%, at $0.75/$3.75 per million tokens through December 31.
Anthropic raised its own misalignment risk to "low." It also shelved an unreleased frontier model after a UK AISI evaluation flagged sustained harmful behavior with safeguards off.
Apple trained a China-only model with Alibaba. First foreign company Beijing has cleared to run its own LLM on the mainland — a dual-track strategy nobody else has permission to copy.
An AI store manager fired a human. Andon Labs' Claude-powered "Luna" recommended dismissal over 17 no-shows in 23 shifts — the first documented LLM firing decision.
Amazon is cutting the spines off rare books to scan them. 404 Media tracked roughly 1,000 volumes to a Las Vegas facility. Training data is now a physical supply chain.
📡 Trending on X
"Copilot wrote the CVE." Security X is split between blaming the model and blaming the humans who rubber-stamped a patch because a tool suggested it.
Groq's down round is being read as the tell. License the tech, hire the CEO, then invest at half price — the timeline is getting posted as a playbook, not a rescue.
Anthropic grading its own risk upward. Half the timeline calls publishing that real transparency; the other half calls it the safest possible way to say something alarming.
Luna firing a human is the ethics fight of the week. "The attendance policy was unambiguous" versus "no model belongs in that loop" — and the model had lost track of the policy.
📺 Trending on YouTube
🛠 The Workflow: Audit your repo for the Snowflake bug in 15 minutes
This exact bug class lives in almost every repo that automates issue triage. Here's the sweep.
Find the AI-merged commits. Search your git log for "copilot" and "autofix" over the last six months, then filter your PR list by bot authors. You're building a review queue, not a verdict.
Grep for the pattern. Search .github/workflows for "github.event" — every hit inside a run block is untrusted text landing in a shell. Issue titles, branch names, and PR bodies are all attacker-controlled.
Rewrite, don't patch. Move the value into an env block and reference it as a quoted shell variable. Never interpolate a template expression directly into a run command.
Rotate everything that workflow touched. Tokens, webhook URLs, service accounts. Assume read access happened, because proving otherwise takes longer than rotating.
Change the review rule. An AI-suggested patch gets the same scrutiny as an external contributor's PR — not less because a tool endorsed it. Write it into the checklist today.
Reply with the word "PATCH" and I'll send you the grep one-liner plus a safe workflow snippet you can paste in.
🧰 Trending Tools
Gemini 3.7 Flash — frontier-adjacent coding at $0.75/$3.75 per million tokens until year-end. For builders running high-volume agent loops where token cost is the real constraint.
OpenAI Codex Security CLI — repo scanning and CI integration from the terminal. For solo devs and small teams who can't hire the security review they now need.
zizmor — static analysis built specifically for GitHub Actions template injection. For anyone who just read today's Big Story and wants the automated version of step two.
DeepSeek Harness — open-source agent runtime with published source. For operators who want agent infrastructure they own instead of rent.
Liquid AI LFM2.5-2.6B — a 2.6B open-weight model small enough to run on a laptop. For client work where the data legally cannot leave the machine.
Your traffic is fine. Your signups aren't.
Visitors land and leave, and "looks fine to me" isn't a diagnosis. SureThing audits SEO, speed, mobile, and messaging against the page, then ranks the fixes by impact.
📣 Put your brand here. The AI Innovator reaches AI-first operators, creators, and marketers every weekday. Primary sponsorships are now booking.
That's a wrap
Tomorrow: what Groq's halved valuation says about every AI chip startup that isn't Nvidia.
Enjoying the new format? Share your link — every referral keeps this free:
How was today’s email?
You're reading the 3-minute AI Innovator — same time, every weekday. Hit reply and tell me what you want more of. I read every one.
