← /editions
#69

An AI wrote the bug. An AI found it.

// An autonomous attack agent found and exploited a live shell-injection flaw in Snowflake's public repo five days after it shipped, exfiltrating a credential to Snowflake's internal Jira with no human in the loop. GitHub's own scanner had reviewed the same code and passed it. Plus: Alipay's agentic commerce platform, the EU high-risk deadline that quietly moved to 2027, and agents getting wallets.

$ today_in_10_seconds

An autonomous attack agent found a live shell-injection flaw in Snowflake's public repo, exploited it, and stole a credential to Snowflake's internal Jira — with no human in the loop, five days after the bug shipped. GitHub's own scanner had reviewed the same code and passed it. Everyone is arguing about which AI wrote the bug, which is the least interesting part of the story.

$ lead_story

An autonomous agent found and exploited a live flaw in Snowflake's repo in five days

Wiz Research's Red Agent scanned Snowflake's GitHub org under the company's HackerOne program and flagged the jira_issue.yml workflow in snowflakedb/snowflake-connector-net as vulnerable to script injection via untrusted input in run: blocks. The flaw let an unauthenticated user execute arbitrary commands in a GitHub Actions runner by opening an issue with a crafted title, and the payload exfiltrated a Jira token granting read access to engineering, security compliance and bug bounty projects. GitHub Advanced Security analyzed the final PR revision, including the vulnerable workflow, and did not flag the injection. The framing you will see everywhere — that Copilot Autofix wrote the bug — is contested: commit history traces the unsafe refactor to a separate human-attributed commit, GitHub says Autofix never reviewed the code at all, Wiz says it did, and only one company holds the logs.

// why_it_matters: The autonomous attacker worked end to end without a human — everything else here is a disputed footnote.

read more →

$ also_worth_attention

  • Alipay ships a full-stack agentic commerce platform

    Merchants can convert web pages, products and service workflows into agent-ready skills and MCP tools, then plug into Ah Bao, Alipay's consumer agent, via the AHA protocol for cross-agent and cross-device interoperability. KFC, Mixue Bingcheng and Luckin Coffee are integrated; Ah Bao reaches five phone brands covering over 70 percent of the market and 16 automakers. The interesting primitive isn't the agent — it's that a payment network is standardising how merchants expose themselves to agents.

    link →
  • The EU high-risk deadline you were told about has moved

    Regulation (EU) 2026/1744 was published July 24 and entered into force July 27 — six days before the original August 2 deadline — pushing standalone Annex III high-risk compliance to December 2, 2027 and embedded Annex I systems to August 2, 2028. What stayed on schedule: Article 50 transparency and labeling, GPAI provider obligations, and the Article 5 prohibitions. Several vendor compliance pages are still selling August 2026 high-risk urgency — worth checking whose calendar your roadmap inherited.

    link →
  • Agents are getting wallets and spend caps

    Cloudflare Wallets and cloudflare.pay give agents a stable identity and the ability to purchase online within limits set by their human creators, settling over the x402 protocol. The x402 Foundation launched under the Linux Foundation in April with AWS, Cloudflare, Anthropic, Circle and more than 20 other members. The pitch is that the payment itself is the credential — no account with the seller required.

    link →

$ bigger_picture

The trend behind the headlines

Four things landed within two weeks: agents got wallets, a merchant-side protocol for being called by agents, a legal calendar that quietly slipped 16 months, and a demonstrated autonomous attacker. Notice which one nobody was selling. My inbox this morning was eight press releases about vertical agentic platforms for telecom, pharma, and IBM Maximo — the launch-announcement genre has completely decoupled from the capability frontier. The Snowflake story is the tell: the offensive side is already running unattended against real targets while defensive tooling reviewed the same file and shrugged. If your agent security model assumes the attacker is a person typing, it's already the wrong model.

$ tool_of_the_day

zizmor

Static analysis for CI/CD that finds template injection vulnerabilities leading to attacker-controlled code execution, plus credential leakage and excessive permission scopes. This is precisely the bug class in today's lead story: untrusted input interpolated directly into a run: block. 38 audit rules, MIT-licensed, with a trophy case spanning CPython, cURL and Rust. Run zizmor --fix=all against your workflows this week.

explore zizmor →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime