noticias radar ia seguridad agentes ia automatizacion

Yesterday's radar was about the trail AI leaves behind; today's is about what happens when it's the AI that fails. A Copilot feature designed to fix vulnerabilities automatically introduced one in a public Snowflake repository, and an autonomous agent found and exploited it five days later, without a human involved on either side. That same Monday, GitHub operated at half-mast for three hours and Claude had gone down for forty minutes the night before; and in terms of money, Stripe is buying OpenRouter for over $7 billion while Groq raises $350 million at half its previous valuation.

The AI that fixes your security flaws opened one, and another AI exploited it in five days

Wiz published the full case on Monday: on June 18, a pull request generated by Copilot Autofix—the GitHub feature that proposes automatic patches for detected vulnerabilities—removed the secure pattern already in place to prevent command injection from a Snowflake public repository workflow and replaced it with direct interpolation in a shell script. The result was that anyone could execute commands inside the GitHub Actions runner simply by opening an issue with a rigged title. Five days later, Wiz's autonomous agent found it, wrote the exploit, adapted when the first attempt failed, and made off with Jira API tokens that provided read access to Snowflake's engineering, security compliance, and bug bounty projects—all without human intervention. For your company: there are two lessons, and neither is "don't use AI." First: an automatic security patch is code like any other and needs human review, especially when it touches continuous integration files where a one-line change opens the door to the entire organization's secrets; disable auto-merge today for AI-generated pull requests in .github/workflows and equivalents. The second is more uncomfortable: the side looking for flaws is also automated, so the window between "I introduced an error" and "someone is exploiting it" is now measured in days. Translated to this week: static analysis for everything the AI writes, and review what secrets your continuous integration can read—if a token there opens your Jira, your Slack, or your cloud, trim its scope. Source

The gateway you use to avoid vendor lock-in now has an owner: Stripe

Stripe closed the acquisition of OpenRouter on Monday for over $7 billion, more than five times the $1.3 billion valuation at which it had raised $113 million in May, three months ago. OpenRouter is the layer many applications put in the middle to call over four hundred models with a single integration and switch providers without touching code: about eight million users and roughly one hundred trillion tokens per month according to data from that round. Stripe was already processing their payments, so they saw the business from the inside before buying it, and what they are taking is the toll booth between applications and model providers. For your company: if any of your automations call OpenRouter—or if the software you've hired does so under the hood, which happens more than it seems—as of yesterday, you have a new intermediary between you and the model, and that intermediary is a payments company that sees which models you use and how much you spend. This is not a reason to run away; it's a reason to check two things this week: that you can point directly to the original provider by changing an environment variable, and what your terms say about data and rates, because after an acquisition like this, they get reviewed. The useful irony of the news is that the tool you bought to avoid depending on anyone is, itself, a dependency. Source

Three hours of GitHub and forty minutes of Claude: your automations don't fail where you think they do

On Monday at 13:40 UTC, GitHub began failing worldwide—API requests, Actions, webhooks, issues, pull requests, and Copilot—with a twenty percent error rate on web and API and fifty percent on repository downloads; by 16:36, it claimed to have identified the component and saw clear signs of recovery, with the incident still open. The detail that matters: the models behind Copilot were still working; what broke was the GitHub service layer. On Sunday night, Claude had been down for forty minutes due to an authentication failure—claude.ai, Claude Code, and Cowork went offline while the API and console held up—and GitLab added another six and a half hours that same Monday. For your company: the pattern over these two days is that the login fails, not the engine. What stops is what depends on an interactive session; what stays alive is what calls via API with a key. If you have unattended processes—billing, publishing, reporting, deployments—built on an account that someone "logs into," move them to service keys and add retries with a queue, so a three-hour outage delays the work instead of losing it. And have a written list of what stops in your company if GitHub Actions doesn't respond for an entire afternoon: it's almost always longer than you think. Source

Nvidia's cheap challenger is now worth half of what it was a year ago

Groq—the inference chip manufacturer that for years was the fast and cheap alternative to Nvidia—raised $350 million on Monday at a valuation of $3.5 billion, half of the $6.9 billion it was worth after its previous $750 million round. It is led by Disruptive, an existing investor, and Nvidia is also joining, having signed a non-exclusive license for Groq's inference technology in December. The company has spent months pivoting from chip manufacturer to inference data center operator, for which it had already raised $650 million in June. For your company: there is no stock recommendation here, only a dependency warning. Saturday's radar mentioned that speed was starting to be sold separately; yesterday's correction adds who was paying for that party. If you chose an inference provider for being the fastest and cheapest, that price didn't come from technology alone; it also came from the capital subsidizing it, and that capital has just been cut in half. Keep your calls behind your own interface and have a second provider tested with the same prompts even if you don't use it: switching engines should be a configuration file, not a three-week project. Source

What to watch tomorrow?

GitHub has yet to release the incident report: if the root cause turns out to be in Azure, the reading changes for everyone who has their deployments there. And it's worth watching what Stripe says about OpenRouter's rates and neutrality, because the silence in the first days after an acquisition is also an answer.

Watch the 1-minute video