Deepfakes, patched switches that stayed open, and AI that writes exploits
Four stories about evidence gaps: from political ads to kernel exploits, and new auditing guidelines from India.
- 01
Deepfake attack ads without disclosure in North Carolina
Republican campaign used AI to create false images of opponent in drag clubs and defacing property, published as political ads without saying they were AI-generated.
Why it matters for evidence: Article 50.4 requires disclosure when AI creates deepfakes. Political ad rules add sponsor and targeting transparency, with records kept seven years.
Source: OECD AI Incidents Monitor
- 02
Zyxel patch fixed the wrong bug, original exploit still works
Zyxel's patch for GS1900 switches closed CVE-2026-7273 but left the exploited CVE-2026-107441 open. Binarly found devices remain vulnerable to the original attack.
Why it matters for evidence: CRA's 24-hour reporting clock starts when you discover active exploitation. A log of what you patched, when, and what you knew matters in the timeline.
Source: Help Net SecurityOur explainer →
- 03
OpenAI model wrote 948 root exploits in one run
OpenAI's upcoming model automatically generated 948 working Linux kernel exploits and patches against the latest kernel in a single run, showing AI can systematically find and weaponise vulnerabilities at scale.
Source: Help Net SecurityOur explainer →
- 04
India publishes AI security auditing guidelines
CERT-In's new guidelines establish a risk-based framework for auditing governance, security, privacy, fairness and transparency of AI systems throughout their lifecycle.
Source: cert-in.org.in