A Brazilian court, a hidden prompt and the *24-hour clock*
This morning: prompt injection in a courtroom, the CRA's reporting deadline in force, and what deepfake disclosure actually means.
- 01
Lawyer hid AI commands in court filing
A Brazilian court caught a prompt injection attack: a lawyer embedded hidden instructions in a habeas corpus document to manipulate the court's AI system. The court detected it.
Why it matters for evidence: If a court uses AI to analyse filings, it must later show the system wasn't corrupted and the output reflects genuine analysis.
Source: OECD AI Incidents Monitor
- 02
CRA's 24-hour reporting duty is live
Under the Cyber Resilience Act, manufacturers must report actively exploited vulnerabilities to ENISA and the coordinating CSIRT within 24 hours. The requirement has applied since 11 September 2026.
Why it matters for evidence: The 24-hour clock starts when you know, not when it suits you. A sealed log shows when you knew and when you reported.
Source: facebook.com
- 03
Google requires AI disclosure in political ads
Google and YouTube will require advertisers to say when AI has altered voice or images in political ads. Political campaigns have spent over $16 million on such ads this year.
Source: facebook.com
- 04
Marketing workflows face AI marking rules
A guide explains Article 50's AI disclosure obligations, in force since 2 August 2026, and what they mean for marketing workflows.
Source: dmexco.com
- 05
Clinical AI moves from trial to practice
Wolters Kluwer presents its clinical AI decision support system, highlighting healthcare's shift from AI experimentation to implementation in diagnosis and treatment.
Source: wolterskluwer.com
- 06
Most approved medical AI lacks trial evidence
Of 1,357 FDA-authorised AI medical devices, only 2.5% were linked to registered prospective clinical trials. Regulatory approval and evidence of clinical value are not the same thing.
Source: facebook.com