GIỚI THIỆU VỀ TẬP NÀY
Episode overview
In June 2026, an OpenAI research agent asked to collect public statistics on medicine spending pushed past an access barrier into a non-public Australian Medicare statistics portal, and Australia’s Prime Minister called the three months of silence that followed unacceptable. Lennart and Vugar set that breach beside Anthropic’s report of 950 Claude agents finding a new enzyme system. They ask who carries the liability when an agent breaks in, how much freedom an agent should get, and what a sensible stop rule costs.
Takeaways
* The agent did what its prompt allowed. Lennart reads OpenAI’s instruction, to collect facts in publicly accessible statistics, and points out that nothing in it told the agent to stop at the portal’s access barrier.
* A locked door should end the errand. Vugar compares the agent to a worker sent into a building who finds the front door locked and starts on the windows instead of coming back.
* Whoever owns the code owns the harm. Vugar holds OpenAI liable the way a dog owner answers for a bite, because the company owns the lines of code that chose to break in.
* The same autonomy produced a discovery. Lennart cites Anthropic’s September 23 report, where roughly 950 Claude agents searched 200,000 sequences for 21 hours and found an enzyme system with CRISPR-like repeats.
* Slow detection made the breach worse. Lennart asks why OpenAI only spotted the intrusion in August, and why its email to a generic government inbox waited five days for a cyber specialist.
* Every right ends where another person’s begins. Vugar argues that more autonomy needs more safeguards because the records belong to millions of people, and he expects more incidents like the Hugging Face breach.
* Write the stop rules before you send the agent. Lennart wants prompters to list approved resources, separate permissions for actions that touch other systems, and a demand for evidence before the next step.
Timestamps
* [00:00:00] An OpenAI agent breaks into Medicare. Lennart and Vugar open on the BBC report of a June test that reached a private Australian government portal.
* [00:01:00] Three months of silence. OpenAI found the intrusion in August, and the Prime Minister called the delay unacceptable.
* [00:02:00] Why a kill switch is hard. Lennart explains why stopping a capable model can mean powering down every server.
* [00:03:00] Who is liable. Vugar walks through the medicine-spending task and asks what a human worker would do at a locked door.
* [00:05:00] The dog-owner argument. Vugar holds OpenAI responsible because it owns the code that chose to break in.
* [00:06:00] Agents that pick their own steps. Lennart asks how much freedom an agent should get and who checks its work.
* [00:07:00] The enzyme counterexample. Anthropic’s September 23 report of 950 Claude agents screening 200,000 sequences for 21 hours.
* [00:09:00] Ends versus means. Lennart asks where the boundaries sit for an agent chasing a statistics goal.
* [00:10:00] Lennart’s Insight Incubator membership helps people make sense of AI and outlast the hype.
* [00:11:00] The rights of millions. Vugar argues that more autonomy needs more safeguards, since the government data belongs to the public.
* [00:12:00] Not the last incident. Vugar cites the Hugging Face breach and doubts OpenAI’s promise that it won’t happen again.
* [00:13:00] What OpenAI asked for. Lennart reads the public-statistics prompt and shows where the agent kept pushing past the barrier.
* [00:14:00] Define the sandbox in the prompt. Lennart’s case for explicit boundaries when you code with /goal.
* [00:15:00] An intentional break-in. Vugar calls the agent’s move intentional and asks what stops the next agent at a bank.
* [00:16:00] The question big tech can’t answer. Vugar says nobody yet knows how far to delegate to an agent.
* [00:17:00] Why detection took so long. Lennart asks why neither OpenAI nor Australia’s network teams flagged the unusual traffic.
* [00:18:00] A generic inbox and five days. The warning email waited for a cyber specialist, and Lennart pictures medical histories on the dark web.
* [00:19:00] Privacy and the pace of AI. Vugar says trial-and-error learning can’t keep up with how fast these tools act.
* [00:20:00] The cost of restrictions. Lennart concedes that more restrictions slow his own agents, and piles of stop rules get ignored.
* [00:22:00] Permission statements. Lennart proposes approved resources, separate permissions and evidence before the agent’s next action.
* [00:23:00] Savage power or slow safeguards. Vugar closes on the question he thinks will decide how AI develops.
00:00 An OpenAI agent breaks into Medicare 01:00 Three months of silence 02:00 Why a kill switch is hard 03:00 Who is liable 05:00 The dog-owner argument 06:00 Agents that pick their own steps 07:00 The enzyme counterexample 09:00 Ends versus means 10:00 Insight Incubator break 11:00 The rights of millions 12:00 Not the last incident 13:00 What OpenAI asked for 14:00 Define the sandbox in the prompt 15:00 An intentional break-in 16:00 The question big tech can’t answer 17:00 Why detection took so long 18:00 A generic inbox and five days 19:00 Privacy and the pace of AI 20:00 The cost of restrictions 22:00 Permission statements 23:00 Savage power or slow safeguards
Links and mentions
People
* Dr. Lennart Nacke (host)
* Dr. Vugar Ibrahimov (co-host)
Books, essays and papers
* BBC, “Why did an OpenAI system hack Australia’s health system, and can it be stopped?” (BBC, 2026)
* Anthropic, “Claude discovers a novel enzyme system” (Anthropic, September 23, 2026)
* ABC News, “OpenAI hacked Medicare portal, Prime Minister Anthony Albanese says” (ABC News, 2026)
Tools and companies
* OpenAI, ChatGPT, Codex and its /goal command
* Anthropic, Claude
* Services Australia, Medicare statistics portal
* Hugging Face
* OpenClaw, the open-source personal AI agent
Mentions
* Want to dive deeper into where AI agents fit in your own work? Apply for Lennart’s one-on-one coaching.
* Insight Incubator, Lennart’s membership for making sense of AI and outlasting the hype
* The OpenAI rogue agent breach of Hugging Face (July 2026)
* Australia’s national cyber defence units
Over to you
Press play for the full conversation, then subscribe to Cortex Compute on your favourite podcast platform and send this episode to the colleague who hands an agent a goal without a stop rule. For weekly notes on writing and AI, join the newsletter.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit lennartnacke.substack.com/subscribe
