Five Stories for Site Operators
A supply chain attack hit major sites this week. Several other stories also touch on trust in tools you run.
The common thread is simple. Each story shows a gap between what a vendor claims and what you can verify. The Optimizely supply chain attack is the most urgent for publishers.
- Optimizely snippets served malware.
- Kubernetes AI conformance is a listing, not proof.
- A memory guard missed a 268 GB process.
- Ollama setups grow into five pieces.
- Google logged an automated AI attack.
Optimizely Snippets Served Malware
An attacker abused a flaw in Optimizely Web Experimentation. According to Dev.to, the attacker wrote a malicious loader into customer Project JavaScript. The official snippet then served it to every visitor.
At least 178 organizations were affected, including CNN, the BBC, HSBC and Equifax. Visitors saw a crypto address swapper and a card skimmer. iPhone and iPad users also faced the Coruna and DarkSword exploit kits.
Optimizely says it fixed the flaw. Still, publishers should act on their own.
- List every third-party script on checkout and login pages.
- Audit who can edit testing and tag tools.
- Use a content security policy to limit script sources.
- Alert on changes to hosted script files.
Marketing tools run with full page access. Treat them like code you ship.
Kubernetes AI Conformance Is Only a Listing
A Dev.to analysis reviews 24 submissions against Kubernetes v1.35. Its point is practical. A listing shows a vendor filed a statement. It does not show the platform fits your workload.
I agree with that distinction. Buyers should write their own acceptance record. Name the workload, the platform and the Kubernetes version. Then attach the test evidence.
A Memory Guard That Trusted the Agent
One developer described a failure in a Dev.to post. A node helper under Claude Code grew from about 10 GB to 268 GB in minutes. Swap hit 76 GB, and macOS ran out of memory.
The guard checked every minute. It killed nine small leftovers and ignored the real culprit. It trusted the agent process.
The lesson is measurable. Set hard limits by resource, not by process reputation. Our earlier AI roundup covered similar agent strains. Heavy local work also favors laptops (paid link) with more memory.
Ollama Is Only One Piece
Another Dev.to author argues that local AI grows quickly. Ollama handles inference on day one. By Friday you add a vector database, Postgres or SQLite, and a glue framework such as LangChain.
Each added piece brings upgrades, backups and failure modes. Small teams should count that upkeep before calling local AI free. The model is the easy part.
Google Logs a Fully Automated AI Attack
A daily digest on Dev.to reports that Google Threat Intelligence Group documented a fully automated exploitation pipeline. It found and exploited vulnerabilities without a human in the loop. The same digest says Microsoft shipped a Windows component that counts which coding agents you launch.
Treat the details as secondhand until you read Google’s own report. Even so, the direction matters. Defenders should shorten patch windows.
Takeaways by Site Size
Small publishers
Remove scripts you do not need. Keep one owner for each tag tool. Back up important data on external drives (paid link).
Larger teams
Add script monitoring and a written acceptance record for each platform. Cap memory for every agent process. Review vendor access every quarter.
Trust needs evidence. A listing, a guard or a vendor fix is only a start.
As an Amazon Associate, TechMogo earns from qualifying purchases.
