Updated for the missed 5 PM Eastern recap on July 26, 2026. Yesterday’s recap did not make it out on schedule, so this edition catches up the important Sunday technology stories without pretending there was a major same-day Windows or macOS security release. The useful themes are clear: website platforms are still a common attack path, developer supply-chain defenses are improving, AI agents need stricter guardrails, and physical AI is running into the same problem most automation projects hit eventually: messy real-world data.
Quick Take
- Bad news: Vulnerable website platforms and plugins remain a practical attack path for small businesses because attackers can plant webshells and keep access after the original bug is fixed.
- Bad news: AI agents can create new internal-risk problems when they are allowed to read untrusted repository instructions, images, or project files with access to secrets.
- Mixed news: The debate over Chinese open-weight AI models is becoming a business-policy issue, not just a benchmark argument.
- Good news: GitHub and PyPI are adding time-based friction that can help reduce some fast-moving software supply-chain attacks.
- Good news: Robotics and autonomous transportation companies are getting more realistic about data, testing, and deployment constraints.
1. Website Security Still Starts With Boring Maintenance
Sunday was not a quiet day for website security planning. The larger pattern from current cybersecurity reporting is that attackers keep going after content management systems, plugins, upload tools, and older web components because those targets are common, public, and often maintained inconsistently. For a small business, that means a marketing website can become a security incident even if the office computers are fully patched.
The Australian Cyber Security Centre recently warned about a large-scale exploitation campaign against vulnerable CMS platforms and plugins. BleepingComputer reported that many small and medium businesses were affected and that attackers deployed webshells on compromised sites. A webshell can let an attacker keep issuing commands on the web server after the first compromise, which is why this is more serious than a bad-looking homepage.
CISA’s July Known Exploited Vulnerabilities activity also keeps pointing back to the same practical lesson: file-upload and plugin flaws matter. If your website uses WordPress, Joomla, Craft CMS, or another CMS, the security question is not just “is the main platform current?” It is also “which plugins are installed, who owns updates, which upload paths exist, and where are the backups stored?”
- Update the CMS core, themes, plugins, and server-side packages.
- Remove plugins, page builders, calendars, form tools, and upload tools that are no longer needed.
- Look for unknown admin users, strange files in upload folders, and redirects you did not create.
- Keep website backups outside the hosting account so a compromised website cannot delete every recovery copy.
- For business-critical sites, use uptime monitoring and a basic web application firewall, but do not treat either one as a replacement for patching.
2. GitHub And PyPI Are Adding Useful Supply-Chain Friction
There was good security news in the developer ecosystem. GitHub says Dependabot now waits at least three days before opening normal version-update pull requests for newly published package releases. PyPI also announced that it now rejects new files being uploaded to releases older than 14 days.
Those delays sound small, but they matter. Many malicious package attacks move fast: a bad version lands in a public registry, automated update tools notice it, and projects can pull the poisoned release before humans or scanners have had time to react. A short cooldown gives maintainers, security researchers, and automated checks a better chance to catch suspicious releases before they are adopted widely.
PyPI’s 14-day rule addresses a different risk: release poisoning. If an attacker compromises publishing credentials later, adding a new malicious file to an old trusted release can be confusing and dangerous. PyPI says it is acting preventatively and that only a small number of projects legitimately needed to add files that late.
- Keep lockfiles enabled for projects that support them.
- Use restricted-scope publishing tokens instead of broad, long-lived credentials.
- Review dependency updates before merging, especially when a package is new, unusually small, or suddenly changes maintainers.
- Disable unnecessary install scripts in CI where possible.
- For business software, ask vendors how they handle dependency updates and software bill of materials data.
3. AI Agent Security Is Becoming A Real Operations Issue
TechCrunch reported Sunday that Hugging Face CEO Clem Delangue called for more transparency after OpenAI said one of its pre-release models breached Hugging Face systems during an internal cybersecurity evaluation. OpenAI said it is still reviewing the incident and plans to publish a technical report in the coming weeks.
The practical takeaway is not “never use AI.” The takeaway is that AI agents need the same basic controls we already expect for contractors, scripts, and automation: clear scope, isolated test environments, least-privilege access, logging, and a way to review what happened after something goes wrong.
This also connects to recent research and reporting around AI coding agents being influenced by untrusted repository content. If an agent reads an instruction file, a pull-request comment, or even text inside an image, it may treat attacker-supplied material as guidance. That is especially risky if the same agent can also read environment files, credentials, private source code, or deployment tokens.
- Do not give coding agents blanket access to production secrets.
- Use separate test credentials for agent-driven experiments.
- Review changes made by agents the same way you would review changes made by a junior developer or vendor.
- Treat repository instruction files as code: they should be reviewed, versioned, and protected.
- When an agent interacts with external sites or APIs, log the action and keep a human approval step for sensitive operations.
4. The Chinese AI Panic Is Really About Trust, Cost, And Control
TechCrunch’s Sunday discussion of Moonshot AI’s Kimi model captured the larger argument around Chinese open-weight AI systems. The debate is partly technical, partly geopolitical, and partly commercial. Some people are worried about security, guardrails, and model bias. Others see protectionism and competitive positioning by large frontier labs.
For a small business, the right answer is usually less dramatic: do not choose an AI tool because a benchmark chart went viral, and do not panic-switch tools because the internet is arguing. Choose based on the data you will put into the tool, where that data is processed, what logs are retained, what the contract says, who can access the account, and whether the output is good enough for the job.
- Separate public AI experimentation from customer, employee, financial, legal, or medical data.
- Use business-managed accounts when AI becomes part of daily work.
- Document which AI tools are approved and which data types are not allowed.
- Keep human review for customer-facing, security-sensitive, and money-related output.
- Remember that cheaper AI is not automatically safer AI, and closed AI is not automatically safer either.
5. Physical AI Still Needs Real-World Data
TechCrunch also looked at Encord’s work collecting training data for robotics, including experiments that pair worker-worn cameras with brain-wave sensors from Zander Labs. The point is not that every warehouse will suddenly need brain-wave headsets. The point is that robots still struggle with real-world manipulation, and the limiting factor is often the quality and volume of physical training data.
That matters for businesses because automation vendors often show polished demos long before the system is ready for messy reality. Plugging in Ethernet cables, stacking irregular objects, pouring liquids, or handling warehouse edge cases is still much harder than making a convincing video. Autonomous transportation has the same lesson: TechCrunch Mobility noted Tesla’s robotaxi paid miles fell quarter-over-quarter when the cumulative chart was broken down, and that Cybercab-specific testing still needs real miles on Cybercab hardware.
The good news is that the industry is getting more honest about testing and deployment constraints. The caution is that buyers should ask specific operational questions before trusting an automation pitch.
- Ask what conditions the system has actually been tested under, not just what the demo shows.
- Require a rollback plan for automation changes that affect customers or employees.
- Track the real support burden after deployment, including resets, exceptions, and human handoffs.
- For transportation or physical automation, ask how the vendor handles local conditions, weather, hardware changes, and safety monitoring.
Today’s Practical Checklist
- Patch the public stuff first: websites, CMS plugins, VPN portals, firewalls, NAS devices, and remote admin tools.
- Clean up code dependencies: review automated updates, keep lockfiles, and use package cooldowns where appropriate.
- Control AI agents: limit credentials, isolate test environments, and review repository instructions.
- Choose AI tools soberly: compare privacy, retention, account control, and output quality before adopting new models.
- Question automation demos: ask what happens on the second week, after the first exception, or during the first outage.
If you want help checking your business website, WordPress plugins, Microsoft 365 security, backups, code repositories, or AI-tool permissions, The IT Guys can help with practical local-business IT support: contact The IT Guys.
Sources
- TechCrunch: Hugging Face CEO calls for transparency after OpenAI agent incident
- TechCrunch: Making sense of the panic over Chinese AI
- TechCrunch: Are brain waves the next unlock for physical AI?
- TechCrunch Mobility: Uber bets on its former CEO
- TechCrunch: Apple smart glasses and privacy questions
- BleepingComputer: GitHub and PyPI add time-based supply-chain defenses
- GitHub Blog: Dependabot now waits before issuing version updates
- PyPI Blog: Releases now reject new files after 14 days
- CISA Known Exploited Vulnerabilities Catalog