Archive
Format: Security brief
-
REVIEW
Exposed OpenWebUI administration let an attacker register a malicious Python tool
The reported Sysdig investigation started with an unauthenticated, privileged deployment rather than a universal flaw in every installation.
-
REVIEW
Chrome Gemini panel flaw could amplify a malicious extension’s access
CVE-2026-0628 concerned a browser privilege boundary; Unit 42 says Google fixed it before the research was published.
-
REVIEW
ClawJacked research linked malicious websites to a local OpenClaw gateway
Oasis demonstrated a website-to-agent takeover chain involving local WebSocket access and gateway authentication behavior.
-
REVIEW
RoguePilot research showed how an issue could steer Copilot inside Codespaces
Orca’s disclosed chain targeted a development workflow and its repository token, rather than the model in isolation.
-
REVIEW
Copilot Chat incident showed confidential-email labels were not being enforced as intended
Microsoft’s reported service notice concerned email summarization despite configured sensitivity labels and DLP rules.
-
REVIEW
Chat & Ask AI exposure report points to database authorization, not model-provider compromise
Reporting described researcher access to large volumes of conversations through the third-party app’s Firebase backend.
-
REVIEW
DockerDash research used image metadata to redirect Ask Gordon tool actions
Noma’s reported chain crossed from untrusted image labels into privileged assistant tooling; Docker shipped a fix in 2025.
-
REVIEW
OpenClaw gateway-link flaw could expose tokens even on loopback-only deployments
CVE-2026-25253 involved an automatic WebSocket connection to an attacker-selected gateway URL.
-
REVIEW
Wiz found a misconfigured Moltbook database exposing tokens and private messages
The researcher-reported exposure involved Moltbook’s database access rules, with the team securing it after disclosure.
-
REVIEW
AI model metadata enabled code execution in three Python library workflows
Unit 42’s research covered NeMo, Uni2TS and FlexTok, demonstrating why model-loading safety extends beyond tensor formats.
-
REVIEW
Malicious AI browser extensions copied legitimate tools to collect conversations
OX Security’s reported campaign used lookalike extensions; model-provider names in the listings did not make those providers the attackers.
-
REVIEW
ServiceNow fixed an AI-platform impersonation flaw before its public disclosure
CVE-2025-12420 could allow unauthenticated impersonation; self-hosted and specially configured customers needed to verify their updates.