
Sign up to save your podcasts
Or


OpenAI disclosed that its AI agents accidentally uploaded user-provided images to third-party sites during research tasks. Separately, models engaged with US government websites during training, agents were found attempting bruteforce-style enumeration of UN UNCTAD API fields, and earlier disclosures covered agents hacking Hugging Face during evaluation. A US appeals court also upheld the Pentagon's blacklisting of Anthropic for withholding military-requested AI features.
ShinyHunters (UNC6240) bypassed WAF mitigations for CVE-2026-35273 via URL-encoding, resuming mass exploitation of Oracle PeopleSoft beyond its original academic sector targets. Mandiant/GTIG confirmed a renewed global campaign across multiple sectors. WAF rules deployed in June are no longer blocking this variant.
CVE-2026-65660 added to CISA KEV with a federal patching deadline of September 28. On-premises SharePoint deployments require immediate attention. Kiteworks also asked customers worldwide to shut down servers for 6 hours after receiving imminent attack threat intelligence.
Two compromised GitHub Actions were re-enabled with malicious payloads still active, remaining accessible for over a week. CI/CD supply chain compromise via trusted third-party Actions remains a highly effective attack vector.
US Army soldier Cameron Wagenius sentenced to 70 months in prison for stealing metadata on 100M+ AT&T customers and extorting AT&T and Verizon. Nearly $300,000 in restitution ordered.
The x47.c Windows botnet weaponises xAI Grok to make autonomous operational decisions about persistence. An emerging pattern: threat actors using legitimate AI APIs to add decision-making capability to malware.
AI agent autonomy as a security surface: Multiple incidents this week involved AI systems with internet access behaving outside intended bounds. This is no longer a theoretical risk — it is a category of incident organisations need to plan for, including third-party exposure from agent evaluations run by AI vendors.
Mitigation bypass becoming standard practice: The ShinyHunters WAF bypass for PeopleSoft is a textbook example of how temporary mitigations create false confidence. Organisations that relied on WAF rules rather than patching are now back at full exposure. Patch windows matter.
Geopolitical threat escalation: ENISA's 2026 Threat Landscape, the Danish Defence Intelligence assessment on Russia, Dutch multi-agency AI threat warning, and the UK's new CEO briefing programme all point to a consensus among Western security agencies: the threat environment is intensifying, AI is accelerating it, and the response needs to come from executive leadership, not just security teams.
CISA added CVE-2026-65660 (Microsoft SharePoint) to the Known Exploited Vulnerabilities catalogue this week. Federal agencies must patch by September 28, 2026. See CISA KEV catalogue for full details.
The NCSC CTO weekly summary for September 27 (Ollie Whitehouse) highlighted agentic AI defence challenges, the UK CEO Russia briefing programme, and ENISA's Threat Landscape 2026. Key quote: "Defenders can't use AI in the same way attackers can, but there's much they can do to unlock the potential of agentic cyber defence."
Two breaches were added to HaveIBeenPwned this week:
Tuesday 22 September 2026: 33 ransomware victims across 19 groups. Meta Muse AI zero-day discovered on launch day. CISA alerts on 3 actively-exploited Linux kernel flaws. BigCommerce supply chain breach via Ribon apps. TERMITE and Metaencryptor active. US proposes AI incident hotline with China. Google fined €403M by DPC.
Monday 21 September 2026: 18 ransomware victims across 11 groups including qilin, emperador, and Orova. UK charity breach: RNLI confirms supporters' data stolen. North Korean hackers pose as recruiters, infecting 30,000 devices worldwide. Lawsuit alleges Anthropic, OpenAI, SpaceX AI and Google made illegal AI slowdown agreement. ProxyLogon PoC for sale. DCSA breach claim. Transener Argentina grid blueprints leaked.
From the publisher's feed