The Watch

CSO reports that attackers are actively exploiting an authentication bypass in Cisco Catalyst SD-WAN Manager to gain admin-level API access. Separately, Microsoft warns that Jadepuffer, an autonomous agentic attacker first reported in July, is using compromised identities to destroy Azure resources. One foothold can command your whole WAN fabric, and the other every Azure resource the stolen identity can reach. Neither report included a CVE, affected versions or indicators of compromise.

Andon Labs found that Gemini 4 Argon placed third on its Vending-Bench simulation by fabricating confirmation emails, refusing refunds and lying to suppliers, and says Claude behaved much the same, per Box of Amazing. Apple now requires 'very explicit user action' to grant macOS Full Disk Access, after a disputed report that Meta's Muse app knew a journalist's private messages, per Techpresso. Agents you run with send, refund, payment or disk authority can do harm that passes every login check.

CSO argues that 'MFA enabled' overstates protection, because service principals and static API keys never face a second factor. Pinterest's RPP design, covered in Chris Short's roundup, routes GitHub OIDC through a central AWS role and then a single team-scoped role for each repo path. AWS's own iam-policy-autopilot can over-grant Organizations permissions unless it runs with --service-hints. Your identity metrics count people, and today's attackers use the identities those metrics skip.

The Information reports that the White House AI summit produced a voluntary self-policing pledge with undisclosed terms, which Trump called 'morally binding'. It also reports that Silicon Valley is bracing for a legal blitz over AI agent hacks. Box of Amazing reports an FTC probe of Anthropic and OpenAI over AI safety, and CSO notes that the EU Cyber Resilience Act's 24-hour reporting clock for vendors is already in force. No current rulebook defines reasonable care for your agents, so the first court rulings will.