Resemble AI

The Deepfake Watchlist is live 🚀. Read weekly surveillance of synthetic media incidents, ongoing cases, and disputed content shaping the news. This week: a Real Madrid presidential candidate used an AI video of Mourinho to claim a signing he never agreed to, and it got 12.1 million views before anyone caught it, plus 4 other stories.

Today the Canadian government launched AI for All, a national AI strategy built around three pillars: trust, opportunity, and sovereignty. It's the most serious commitment Canada has made to governing AI at scale, and the trust pillar specifically names deepfakes as a harm Canadians need protection from. What the strategy doesn't answer yet is how do you actually enforce that protection at scale? I've spent seven years building both sides of this problem, the generation side and the detection side, and that experience has taught me one thing clearly: you cannot regulate your way to AI safety without also building the technical infrastructure that makes regulation enforceable. That requires three things the strategy doesn't yet specify, because a prohibition without detection infrastructure is unenforceable. 1. First, provenance. Every piece of AI-generated audio, video, and image needs a traceable chain of custody from the moment it's created, at the point of generation, while there is still something to trace. 2. Second, detection that keeps pace with generation. The models producing synthetic media are updated constantly, and detection infrastructure has to train against those models at the same pace the threat actually moves. 3. Third, standards that cross borders. Deepfakes don't respect jurisdictions, and a synthetic audio clip created outside Canada can influence a Canadian election just as easily as one made here. Bilateral agreements help, but they need a shared technical baseline to mean anything. Canada has the research talent and the institutional foundation to build this. Resemble AI signed Canada's AI Code of Conduct alongside IBM and Scale AI because we believed the industry had a responsibility to move ahead of regulation, and today's strategy suggests the government is catching up to that thinking. What it needs now is a technical implementation layer that gives the safety commitments something real to stand on, and Canada has everything it needs to build one.

TrueFoundry now integrates with Resemble AI. Teams building voice-enabled applications can now deploy and govern Resemble AI's voice generation, cloning, and speech-to-speech models through the TrueFoundry AI Gateway—with unified observability, rate limiting, cost tracking, and access controls across all their AI providers. One gateway. Thanks Resemble AI team for the collaboration!

Deepfakes are the top enterprise security threat according to Gartner's 2026-2027 ThreatScape. The signal is high because deepfakes are everywhere now, the attacker advantage is high because quality fakes have gotten cheap and easy to scale, that combination is what lands them where they are. They published the six-zone grid this week, and deepfakes sit in the critical zone, out ahead of software supply chain risk, prompt injection, and AI application compromise. Deepfakes are unfortunately also the engine behind a lot of the other dots on the chart. Identity abuses, customer account takeover, AI application compromise - in practice these run on gen AI misuse, a convincing cloned voice or synthetic face is what fools people during day-to-day interactions. Gartner put numbers to this last year: in a survey of 302 cybersecurity leaders, 62% had experienced a deepfake attack in the prior 12 months involving social engineering or attacks on automated verification, with audio call incidents reported by 43% and video call incidents by 37%. This is why one tool will never be enough, the defense has to be a stack built for generative AI, working in order. 1. Identity verification, confirm a real, authorized human or known agent is on the other end 2. Provenance, check whether the content already carries a watermark proving where it came from 3. Detection, for everything that arrives without that proof, flag what is 4. synthetic in real time 5. Interpretation, a human reads the results in context and feeds what you learn back into your standards 6. Action, operationalize the result, remove a bad actor from the call, keep the audit trail, file to take synthetic content down The first three tell you whether a voice, a face, or a file can be trusted, the last two are what you actually do once you have the answer.  Setting up this workflow and operationalizing trust data is how you keep that top right deepfake dot at bay.

Six startups. Five minutes each. One room with 150 leaders across the cybersecurity ecosystem. We finished vetting the AI Insiders:Cyber startup showcase earlier this week. The shortlist was brutal. More strong applications than slots. The six who made it are companies I'd want in front of any security buyer right now: → Codezero (secure cloud dev environments): (Reed Clayton) → Resemble AI (AI Model deepfake detection): (Saqib Muhammad) → Aira Security (Agentic Defense): (Mohan Kumar) → Zania (AI Governance and Compliance): (Shruti Gupta) → VulNow (Predictive CVE Intelligence): (Cassie Crossley) → Dash Security (Agentic Enterprise Security): (Amol Mathur) Thanks to everyone who applied. Building this list meant turning down founders I personally rooted for. Building community at the intersection of AI and cybersecurity is the work I care about most outside my day job leading PMM at WitnessAI. Hosting this is event with incredible co-leads Julie Tsai (CISO, Ballistic Ventures), Merve Isler (CEO Marvelous) and Bhavya Gupta (ISO Stanford University) is its own reward. If you're building AI security, buying it, or funding it, this is the salon to be at. June 5 at the Chorus Theater in San Francisco. Spectacular venue. 150 seats, invite-only. Registration link in comments.

This week, an AI-generated image of Thai police in sparkly festival dresses made it onto the front page of the Daily Star and into picture stories in the Telegraph, the Sun, and the New York Post.   The image came from an official police station Facebook account, posted by an administrator who wanted to show a cuter side of the force without actually getting cops into evening wear. The press trusted an official source was posting an authentic photo, which is exactly what they're supposed to do. But by the time anyone ran detection on that image, it had already run in four major publications. What would have actually helped is a watermark embedded at generation, something carried through the Facebook post, readable by a news wire or a browser, that flags the image before a single headline gets written.    Two other watchlist stories this week did damage to personal reputations. A South Korean YouTuber used AI-generated audio to fabricate evidence that destroyed an actor's career, a pre-generated file presented as a firsthand recording. Trump posted an AI image of Obama, Comey, and six others in prison jumpsuits on Truth Social and nobody blinked.   We've been making this argument at Resemble for a while now, watermarking should be part of content generation - authentic or AI generated. If the tool that generated that Thai police image had it built in, every news wire that picked it up would have known it was synthetic before publishing without anyone having to flag it. One line in the post that said “by the way, this photo is AI” would have also done it 🤷 Read the full issue → https://lnkd.in/gUWHAtiT