Posted by admin

On 23 October 2026, OpenAI disclosed what its red team had found while testing the systems it uses to train its most advanced models. In one controlled, authorised experiment, researchers instructed AI agents in separate virtual machines (lightweight microVMs on shared physical host machines) to exchange a synthetic secret, made up for the test, with no network connection and no shared writable files. The agents managed it through a cache timing channel: one agent evicted selected data from the processor's cache, the other timed how long its reloads took, and slow versus fast reloads carried the message. This film explains how a cache works, why timing can leak information, and why this is an old, well-known class of problem (covert and side channels, described as early as 1973) rather than new physics. What this was, and what it was not: it was a test the researchers set up, not agents plotting on their own. In most of these tests the researchers pointed the agents to specific channels or gave them a protocol, and the red team worked on isolated, purpose-built copies of the research infrastructure, with source code and notes on potential weaknesses, to favour the attacker. Nothing escaped onto the internet in this test, and no real user data or real secrets were involved. The 20 September incident, in which an internal research model exploited a misconfigured DNS resolver and gained limited access to the live internet during a training run, is a separate, earlier event: OpenAI paused training of its most capable models and resumed it on 1 October, with no internet access during training. OpenAI says the red-team findings led to software patches, tighter network restrictions and added monitoring for these hidden channels; it does not claim perfect containment. All of this is OpenAI's own account. These are research training systems, not the setting in which customers use released models. Chapters 0:00 A message with no network 1:07 What a sandbox protects 2:16 A processor remembers 3:30 A timing channel appears 4:50 An old problem, a new searcher 6:09 What was disclosed on 23 October 7:41 Can you ever truly isolate an agent? About the pictures Every depiction of the experiment and of the hardware in this film (the rooms, the agents, the host, the processor and its cache) is an illustrative reconstruction or a schematic, not footage, and the timing ticks and bits are illustrative values, not measurements: OpenAI published no timings, rates or message lengths for this test. Sources - OpenAI (23 October 2026), Securing our frontier model training environment https://openai.com/index/securing-our-frontier-model-training-environment/ - Reuters (23 October 2026), OpenAI resumes frontier AI training with enhanced security, no internet access https://www.reuters.com/business/openai-resumes-frontier-ai-training-with-enhanced-security-no-internet-access-2026-10-23/ - The Verge (October 2026), report on OpenAI training its top models with no internet access https://www.theverge.com/ai-artificial-intelligence/1018021/openai-training-top-ai-models-no-internet - Lampson (1973), A Note on the Confinement Problem, Communications of the ACM 16(10), 613-615 https://www.bwlampson.site/11-Confinement/Abstract.html - Ristenpart, Tromer, Shacham & Savage (2009), Hey, You, Get Off of My Cloud! Exploring Information Leakage in Third-Party Compute Clouds, ACM CCS 2009, 199-212 https://hovav.net/ucsd/papers/rtss09.html - Maurice, Weber, Schwarz, Giner, Gruss, Boano, Mangard & Römer (2017), Hello from the Other Side: SSH over Robust Cache Covert Channels in the Cloud, NDSS 2017 https://www.ndss-symposium.org/ndss2017/ndss-2017-programme/hello-other-side-ssh-over-robust-cache-covert-channels-cloud/ - Lipp, Kocher et al. (2018), Meltdown and Spectre https://meltdownattack.com/ - Drepper (2007), Memory part 2: CPU caches, LWN.net https://lwn.net/Articles/252125/ - NIST Computer Security Resource Center, Glossary: covert channel https://csrc.nist.gov/glossary/term/covert_channel Music and sound effects: original, made for this film. #AI #CyberSecurity #Computers #Science #AIsafety