Blog
RSS-
Cover: Reversible Privacy for AI Coding Agents
How Cover replaces private values with deterministic fakes before an AI request, then restores the originals locally without breaking agent context.
- cover
- ai-agent-security
- privacy
- llm-security
- pseudonymization
- codex
-
CVE-2026-66804: CrossDevice Frame Server LPE
How a user-writable CrossDevice COM server path let a current-user Media Foundation virtual camera load native code through Windows Camera Frame Server.
- windows
- vulnerability-research
- local-privilege-escalation
- media-foundation
- com
- crossdevice
- cve-2026-66804
-
CVE-2026-49176 Exploit Development: WalletService to SYSTEM
Exploit development for the Windows WalletService vulnerability, from caller-controlled known-folder resolution to a persisted ESE callback and an interactive SYSTEM shell.
- windows
- vulnerability-research
- cve
- local-privilege-escalation
- exploit-development
- walletservice
-
Windows AppResolver LPE: From AppContainer to SYSTEM
Exploit development for a Windows AppResolver authorization issue fixed in July 2026, from a zero-capability AppContainer to an interactive SYSTEM shell.
- windows
- vulnerability-research
- local-privilege-escalation
- exploit-development
- appcontainer
- uac-bypass
-
CVE-2026-58635 Exploit Development: From a Braille Table to SYSTEM shell
Exploit development for the Windows Narrator Braille vulnerability, from an unauthenticated BrlAPI parameter write to LocalService execution and an optional SYSTEM shell.
- windows
- vulnerability-research
- cve
- local-privilege-escalation
- exploit-development
- accessibility-security
-
CVE-2026-52860: Vim Arbitrary Code Execution
Root cause and exploit primitive for GHSA-65p9-mwwx-7468 / CVE-2026-52860, a Vim Python omni-completion arbitrary code execution bug fixed in 9.2.0597.
- vim
- vulnerability-research
- cve
- exploit-primitive
- editor-security
-
Finding Gaps in Gemini CLI Security Boundaries
Two rejected Google VRP reports and one design concern from my Gemini CLI testing: a blocked-command bypass, an MCP startup policy bypass, and an MCP permission-scope issue.
- gemini-cli
- ai-agent-security
- cli-security
- mcp
- sandboxing
- rejected-finding