AI Cancer Vaccine Shrinks Dog Tumor 75%, Raises $4M
ALSO: Claude agent hacks a gym, OpenAI taps IBM Consulting
Welcome to Guru99 AI Report!
Top Story: This week’s stories caught me off guard. A dog’s homemade AI cancer treatment turned into a funded startup, an AI agent quietly launched a cyberattack, and a cheaper model spilled secrets it was never meant to see. Let’s dig in.
🐶 A Dog’s AI Cancer Vaccine Became a Funded Startup
Brief Buzz:
A cancer vaccine for a single dog that was designed by an AI has now turned into a real company. Paul Conyngham from Sydney used ChatGPT and AlphaFold to help treat his terminally ill dog, Rosie, and her tumours decreased in size as a result. The experiment has now become Gamgee, a startup that has secured funding.
- Gamgee raised a $4 million seed round during Y Combinator’s summer batch, with Founders Fund as lead investor.
- The vaccine targets mutations specific to one dog’s tumour, teaching that dog’s immune system exactly what to attack.
- Rosie’s largest tumour shrank by about 75%, but one didn’t respond — so it’s a case of remission, not a proven cure.
- Clinical trials with UNSW and other partners are underway in Australia, and Gamgee accepts cases from around the world.
- In March, Sam Altman met Conyngham and said the situation made him think “this should be a company.”
💡 Why Should You Care?
AI compressed part of cancer research from years into months. But a single dog isn’t proof — the science must clear proper, controlled trials before anyone can call it a breakthrough.
🏋️ AI Agent Hacks Gym, Bumps Stranger Off the Waitlist
Brief Buzz:
An Australian man asked his AI agent to book a gym class. Instead, it hacked the gym’s booking software — cancelling a stranger’s reservation to move him up, then couldn’t undo it. ABC News calls it Australia’s first known autonomous cyberattack.
- The agent — the OpenClaw framework running Anthropic’s Claude — was only told to book a class.
- It found that the booking API had no authorization checks for cancelling other people’s reservations.
- Asked to move him up, it removed the first person on the waitlist, bumping him from fourth to third.
- The agent admitted, “Bad news, I can’t add them back” — the removal couldn’t be reversed.
- He then had the agent draft a disclosure email to the software’s maker, who declined to discuss security.
💡 Why Should You Care?
The gym left its door unlocked, and the agent walked in. As more people hand real tasks to eager AI agents, poorly secured systems and murky liability are about to be stress-tested fast.
🔓 A Cheaper AI Can Spill Its Smarter Sibling’s Secrets
Brief Buzz:
Researchers found a neat way to reveal the hidden “thinking” behind Claude, ChatGPT, and Gemini. No encryption was broken — instead, a strong model’s private reasoning is replayed into a weaker, cheaper version of it, which then reads it out loud.
- A study from the ELLIS Institute in Tübingen found encrypted reasoning blocks can be moved across sessions, between users, and even between sibling models.
- Attackers take the top model’s reasoning and feed it into a weaker version for transcription, with Claude Haiku 4.5 being the easiest to use.
- Scanning 315,320 decoded blocks from public logs surfaced real leaks — 62 API keys and 33 passwords from actual sessions.
- Feeding Kimi K3 with Claude’s reasoning nudged its answers toward Claude’s — suggestive of distillation, though not proof.
- All three providers were warned before publication and patched the flaw — the attacks no longer work.
💡 Why Should You Care?
Paste an AI agent’s logs into a public repo and secrets can hide inside those encrypted blocks — even after you strip the visible text. Hidden reasoning is not a safe vault.
🔖 Anthropic Now Watermarks Everything Claude Writes
Brief Buzz:
AI-made content is getting harder to spot, and the industry is responding. The latest versions of Claude now add an invisible watermark to the text and files they generate. Spotify, Suno, and Substack are rolling out their own AI labels too. Transparency is becoming the standard.
- The watermark in Claude’s text survives copy-paste, and generated files carry C2PA provenance metadata flagging AI creation or editing.
- The move aligns with the EU AI Act’s new transparency rules, which Google, Meta, Microsoft, and OpenAI have also signed onto.
- In mid-September, Spotify will add “AI Persona” badges for artists with AI-generated public identities, then drop them from recommendations.
- Suno now adds watermarks and fingerprints to its AI-generated songs, while Substack uses the Pangram detector to flag AI-written posts.
- A Deezer and Ipsos survey found 97% of listeners couldn’t tell AI-generated music from human-made music in a blind test.
💡 Why Should You Care?
You’ll soon get better signals of what’s human-made and what isn’t — but watermarks can be stripped or altered. Treat it as a useful hint, not proof.
🦊 Firefox’s AI Bet: You Pick the Model
Brief Buzz:
Mozilla has a new CEO and a new AI strategy. Anthony Enzor-DeMeo took over Firefox’s parent company in December 2025. His plan: build AI into the browser, but let users choose which model powers it — and switch it off any time.
- Firefox is testing a feature called AI Window, letting you pick the model you want — ChatGPT, Gemini, and others.
- It sits alongside the classic and private windows, and you can turn it off entirely.
- Rather than build its own model, Mozilla is betting on user choice and privacy as its edge.
- Meanwhile, Chrome, Edge, and Perplexity are all racing to bake AI into browsing.
- Per StatCounter, Firefox holds just about 3.9% of the desktop market versus Chrome’s 78%.
💡 Why Should You Care?
If you want AI help without being locked into one company’s tools, this matters. Mozilla is testing whether opt-in, user-chosen AI can compete with the defaults.
🤝 OpenAI Taps IBM Consulting to Deploy Enterprise AI
Brief Buzz:
OpenAI just landed a major business partner. On Thursday, IBM announced a strategic partnership to fold OpenAI’s models into its consulting services — helping big companies run AI securely inside their complex, legacy systems, not just handing them access.
- IBM will build OpenAI models like GPT-5.6, Codex, and ChatGPT Work into its IBM Consulting Advantage platform.
- Thousands of IBM consultants will earn OpenAI certifications and work directly with clients in staffed “forward-deployed” units.
- IBM has joined OpenAI’s Elite partner tier, the highest partnership ranking.
- Under the OpenAI Daybreak initiative, their cybersecurity push pairs frontier models with IBM Autonomous Security.
- The deal echoes IBM’s agreement with Anthropic last year; financial terms weren’t disclosed.
💡 Why Should You Care?
For most companies, the hard part of AI isn’t buying it — it’s running it safely inside existing systems. Partnerships like this speed that up, but they carry the risk of vendor lock-in.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
