08/17/2026
Five stories I covered this week, and they all failed the same way: the safety signal was still on screen long after the thing behind it stopped working.
Trellix pulled a set of AI attack tools off criminal forums. One of them, APEX AI, takes a target domain and returns a step-by-step ransomware deployment plan. Indirect prompt injection kits are going for $150 a month. A Trellix researcher put it plainly: someone who wouldn't know where to start on a pen test can now buy a prioritized attack plan. The skill floor didn't erode, it got sold off.
https://www.linkedin.com/posts/kamalathedesigner_researchers-find-ai-powered-hacking-tools-activity-7494152219517382656-LZ9U
Anthropic's Claude watermark comes off for about four cents. Forensic testing across 846 runs found KGW-style marks stripped at a 100% rate after one paraphrase pass, and Google's SynthID at 98.3%. The part that should worry anyone shipping provenance features: a public detection API turns into an evasion oracle. You paraphrase, you check, you repeat until it reads human. Detection you can query for free is detection you can defeat for free.
https://www.linkedin.com/posts/kamalathedesigner_four-cents-strips-claude-watermark-anthropic-activity-7493720110214397952-922v
Unit 42 published three attacks on Google-synced passkeys, and the worst one lifts the security domain master key straight out of Chrome's process memory, then decrypts every passkey in the account. No admin rights. No biometric check. The cryptography behind passkeys is intact. What broke is the device trust model wrapped around it, which is the part users were told to rely on.
https://www.linkedin.com/posts/kamalathedesigner_new-pass-ta-key-attacks-let-malware-hijack-activity-7493306024183898112-Qtja
Spur analyzed 6,038 LG and Samsung smart TV apps and found residential proxy SDKs in 2,058 of them, close to half the LG catalog. The consent flow asks once, buried, and the proxy keeps routing other people's traffic through your home connection after you close the app. That is a dark pattern with a network address attached, and LG only moved to ban the practice after the research went public.
https://www.linkedin.com/posts/kamalathedesigner_nearly-half-of-lg-smart-tv-apps-contain-residential-activity-7492965891404451842-9fTa
A single missing Firestore rule left 181,000 tl;dv meetings queryable by any user, from 80,000 accounts, including government agencies in 23 countries and companies like Salesforce and Cloudflare. Attackers could join roughly 80% of private calls by impersonating the notetaker bot. A researcher flagged it in January and could not get anyone at the company to answer. It was still live in August.
https://www.linkedin.com/posts/kamalathedesigner_inside-the-tldv-flaw-that-exposed-live-government-activity-7492587752186867712-eyq2
Every one of these is a design and policy failure wearing a security costume. The lock icon, the consent tap, the passkey prompt, the little watermark badge: all of them are interface promises, and somebody has to own verifying that the backend still keeps them.
An AI notetaker just proved that convenience and access control are not the same thing. A missing tenant boundary in tl;dv's Firestore database let any logged-in user query meeting records across every customer account: 181,874 meetings tied to 84,312 users across 35,003 email domains. More than 1,0...
