Short items that matter for anyone shipping AI-built apps: research, platform security changes, follow-ups and funding. Each one links to its source. These are not full incident entries. News RSS
Another sign that security budgets are moving to AI tools, MCP servers and the apps employees ship with them. Pluto says it monitors those components and agent activity in real time.
Rein's pitch is guardrails at the moment an agent acts, not after. Total funding is now $35M, per the company and Calcalist.
On 186 real-world tasks, SWE-Agent with Claude 4 Sonnet got 57% functionally correct but only 11.8% secure. Adding vulnerability hints to the prompt barely helped. These are benchmark results, not production breach rates.
Researchers collected 9,041 open-source apps built with Claude Code and Lovable and audited 200 deployed ones, finding 1,186 vulnerabilities, mostly broken access control, injection and auth failures. Better prompting reduced them but did not remove them.
Admins get one view of every project with publish status, security findings and PII findings, plus filters for abandoned and ownerless projects. That targets the shadow-app problem RedAccess described.
The quick scan checks database config, missing RLS policies and authorization gaps in about 10 to 15 seconds, and admins can block publishing on critical findings. Lovable says it does not replace a full security review.
Capsule says invisible instructions in files or cloned repos could make the agent act with the developer's local access and carry over to other projects. Cursor closed the report as Informative, citing its program's prompt-injection exclusion.
Vibe