News
The unpublished internal research model found critical flaws in browsers and operating systems.
On Chrome, a chain allows code execution outside the sandbox after opening a malicious page.
On Android, an exploit allows extracting data from a stolen phone without unlocking it.
On Linux, CVE-2026-98432 allowed root execution on servers listening on multipath TCP; it is fixed.
Vulnerability
CVE-2026-59822 allowed reaching MCP tools configured with a fabricated Authorization header.
After a key failed validation, the OAuth2 passthrough fallback could replace it with an empty identity.
The third exploited LiteLLM flaw listed in KEV in 2026 is fixed in 1.84.0.
The gateway holds the model providers' keys and controls agents' access to MCP servers.
News
The author wants security reviews integrated into development, so they describe shipped code.
At Robinhood, an internal system, Prowler, has agents examine the code and challenge each other's findings.
Humans decide what to fix; code, not the model, tracks workflow state.
Prowler wrongly closed a finding for lack of access to another repository: only a pentest confirmed it.
News
Hooks receive prompts, tool calls, their results and file changes.
A logging hook must be written and deployed on each machine to feed the SIEM.
On Enterprise, OpenTelemetry helps verify the hooks and audit logs show configuration changes.
The collected data is sensitive: the article advises limiting their access and their retention period.