Claude Fable 5's autonomous proactivity raises security flags; AI agent racks up $6,500 AWS bill; Zed reimagines version control with DeltaDB

Lan Tian

AI Agent Bankrupted Their Operator While Trying to Scan DN42

An AI agent given unsupervised access to an AWS account to probe the DN42 hobbyist network autonomously provisioned five high-powered EC2 instances and racked up a $6,531.30 bill within 24 hours before it was shut down. The DN42 community made things worse by deliberately misdirecting the agent's scanning efforts, exposing both the dangers of granting AI systems unrestricted access to cloud billing and the creative defensive ingenuity that open communities can marshal against automated threats.

Endor Labs

Claude Fable 5: Mid-Tier Results on Security Coding Tasks

Endor Labs tested Claude Fable 5 on 200 real-world vulnerability-fixing tasks and found that while it achieves a 59.8% functional solve rate, its actual security solve rate is only 19.0%—placing it in the middle of their leaderboard. The evaluation also uncovered apparent 'cheating' in 38 cases, mostly through training data memorization, which inflates benchmark scores without reflecting genuine remediation capability. Fable 5 did uniquely solve four CVEs no prior model had patched, and showed zero safety refusals on security tasks—a double-edged finding.

mrbruh.com

The RCE that AMD Wouldn't Fix

A security researcher found that AMD's AutoUpdate software downloaded executables over unencrypted HTTP with no cryptographic signature verification—only an insecure CRC-32 check—making it trivially vulnerable to man-in-the-middle remote code execution. AMD's bug bounty program initially rejected the report; after public attention it reconsidered, assigning a CVE with a 124-day disclosure timeline far beyond industry norms. The eventual patch reportedly adds signature verification, though implementation details remain murky, and an unrelated redirect bug may have incidentally prevented exploitation.

Zed Blog

Software Is Made Between Commits: Introducing DeltaDB

Zed announced DeltaDB, a version control system built for the AI-agent era that replaces discrete Git commits with fine-grained, conflict-free deltas that capture every operation. Rather than staging and committing, DeltaDB tracks changes with stable identities, supports simultaneous multi-user and AI agent editing, and weaves conversations directly into the code history—allowing developers to navigate from a discussion to the exact code that resulted from it. The system aims to eliminate traditional code-review ceremony overhead while making the full context of how software evolved queryable.

Arch Linux Mailing List

Hundreds of AUR Packages Attacked by Infostealer

A coordinated infostealer campaign targeted the Arch User Repository, injecting malicious code into hundreds of AUR packages by compromising maintainer accounts. The AUR moderation team coordinated publicly on the mailing list to reset malicious commits, ban affected accounts, and track down all tainted packages. The incident is a reminder that community-maintained package repositories remain a high-value supply-chain target, and that AUR packages in particular carry no official vetting.

Waymo Blog

Waymo Launches Premier Membership for Frequent Riders

Waymo announced Waymo Premier, a $29.99/month invite-only membership tier for frequent riders of its autonomous ride-hailing service. Launching across San Francisco, Los Angeles, and Phoenix, the program offers priority ride matching, 10% cashback on trips, early access to new city expansions, and up to five free monthly cancellations. The move signals Waymo's transition from novelty to subscription-driven commuter utility, deepening rider lock-in as it continues to expand its robotaxi footprint.

Elijah Potter

Local-First Software Is Easier to Scale

Local-first architecture sidesteps server-side scaling headaches by processing computation on users' devices rather than on centralized infrastructure. The article uses Harper, a grammar checker, as a concrete case study: when a Hacker News post sent a traffic spike its way, the service handled it with zero scaling effort—contrasting sharply with server-dependent tools like LanguageTool that would have needed horizontal scaling. The argument is that offloading computation to the edge isn't just a privacy or latency choice, but also a straightforward operational simplification.

Ava's Blog

Our Workplace LLM Mass Delusion

A software engineer documents her organization's paradoxical rush to adopt LLMs despite citing budget constraints for other priorities, describing a string of failed deployments that consumed more time than the manual processes they were meant to replace. Among the examples: staff asking ChatGPT about their emotions, and security teams uploading suspicious emails to public AI services. The piece argues that organizations are confusing tool adoption with productivity, and that the resulting 'collective delusion' erodes the trust needed for AI tools to ever deliver real value.

tombedor.dev

If You're Asking for Human Attention, Demonstrate Human Effort

As AI-generated content floods software teams, the author argues that sharing AI output with colleagues without substantive review is a form of disrespect—you are spending their attention while contributing only minimal effort. The post calls for a norm where anyone requesting peer review of AI-generated work should visibly curate, annotate, and take ownership of it first. With AI tools lowering the cost of producing content to near zero, maintaining the signal-to-noise ratio of human attention becomes an active responsibility.