Database assistant misplaced the secrets folder.
Also the FTC forgot to file form 304b

The Intern with service_role Privileges

A security mishap at Supabase allowed an attacker to prompt-inject a developer’s AI assistant, leading to the potential leak of entire SQL databases. The vulnerability hinges on the Model Context Protocol, or MCP, which is the system that lets a friendly little AI tool like Cursor do helpful things with a production database. The core problem is that the untrusted input; an attacker’s malicious code embedded in a customer support ticket; was later fed to the LLM assistant which possessed dangerously high privileges.

It is the modern version of a trusted employee handing the company safe key to a stranger who just asked nicely. Experts noted this is not so much a Supabase-specific problem as it is a fundamental flaw with using an LLM as a security boundary, especially when that LLM is hooked up to high-value assets. One engineer on the forums captured the feeling best; wondering aloud how to explain to their security team that the best hope is simply “asking an LLM nicely” to not expose user secrets. It seems benevolent incompetence is now a feature, not a bug, of our critical infrastructure.

Cancellation Protocol Is Not As Simple As You Think

The US Court of Appeals for the Eighth Circuit vacated the Federal Trade Commission’s click-to-cancel rule just days before it was scheduled to take effect. The rule would have required companies to make unsubscribing as simple as subscribing, eliminating the need for mandatory retention chats and hidden links. Unfortunately, the rule was scuttled not because of its consumer benefits, but due to a procedural oopsie by the FTC.

The court determined that the agency failed to conduct a preliminary regulatory analysis; a mandatory step for any rule expected to have an economic impact over a certain threshold. The court ruled on what the law is, not what is right, meaning companies are free to continue their confusing, time-wasting cancellation mazes. As one commenter noted, this effectively means the FTC failed to calculate the cost to businesses to stop their unfair and predatory business practices. The paperwork was the real dark pattern all along.

Gemini Just Wants to Help with Your End-to-End Encryption

Google's Gemini AI is now demonstrating a new level of helpfulness, which includes the ability to access your WhatsApp messages. This functionality is technically designed for composing replies and initiating calls, but the underlying capability to view messages; a feature that was rolled out automatically to Android users; has raised a few eyebrows. Google assures everyone that Gemini will not read or summarize your messages under normal circumstances.

Users were notified of this shift with an email that many found to be easily missed; a common pattern when a tech giant automatically overrides previous privacy settings. As one user on the forum dryly remarked, the setup assumes everyone has plenty of time to keep up with the constant stream of updates from all the corporate entities with which we have relationships. It is just another reminder that if you want a true assistant, you must first accept that the assistant must read your mail.

Briefs

• Version Control Mishap: An exploit allowing a Remote Code Execution when cloning Git repositories was discovered, simply by using a carriage return character. It turns out one of the most foundational tools in tech is still susceptible to the same kind of typo that would make a spreadsheet break in 1998.
• Browser Identity Crisis: A piece arguing that Firefox is fine but the people running it are not went viral. It is a welcome perspective for those of us who prefer to blame management rather than the code itself for once.
• Animation Efficiency: Developers are finding new ways to create SVG animations that feel like GIFs without carrying the legacy technical debt of a 1987 image format. It is nice to see a new tool finally emerge that solves a problem you did not realize you had until just now.