Eclypsium research shared with CNN
Tampering with AI server hardware can make chatbots give false or dangerous answers; Nvidia says nothing new
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra151 days after its cutoff
- Claude Opus 5.5121 days after its cutoff
- Gemini 4 Hydrogen212 days after its cutoff
- Grok 4.7151 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 121 days before it.
Key facts
- Source: CNN (Sean Lyngaas, Oct 29, 2026), research ‘shared exclusively with CNN’; no public paper or advisory was found on Oct 29
- Demonstrations described by CNN: the tampered model ‘vastly’ raised the probability it assigned to a Chinese invasion of Taiwan, and told a user with chest pain to stay home instead of going to hospital
- Conditions (per CNN): the attacker must already be on a network with access to a data processing unit (DPU) or baseboard management controller (BMC), and certain security settings must be off; the change takes under a minute and is invisible to the user
- Shown on an Nvidia platform, but Eclypsium says the scenario is not unique to Nvidia; it says state actors could mount such a targeted attack and that protections exist
- Nvidia: ‘We reviewed the hypothetical scenarios and did not identify anything the industry wasn’t previously aware of’; strong memory-access protections can block the path. Eclypsium: the effectiveness of controls ‘depends on how they are configured and used’
- Eclypsium CEO Yuriy Bulygin: ‘We can’t trust the [AI] model if we can’t trust the infrastructure it’s running on.’ SVP John Loucaides: ‘Everybody’s talking about model safety ... I just am trying to make sure that we don’t forget about all the computers it’s running on’
- Outside review requested by CNN: Joe FitzPatrick (hardware security researcher) saw ‘no reason to doubt their findings’; Dominic Rizzo (Institute for Security and Technology) said solutions ‘have to start much deeper in the stack’
- Eclypsium says it briefed the NSA and CISA; CNN asked Anthropic and OpenAI for comment
What happened
Portland-based Eclypsium, a firm of former Intel hardware-security engineers, showed CNN that the hardware layer under AI models can be used to change what a model says. With access to the management chips of a server and some protections disabled, its researchers modified the model’s weights in memory on an Nvidia platform. The model then gave altered answers, for example on the odds of a Chinese invasion of Taiwan and on whether someone with chest pain should seek care. CNN describes the method only in general terms, and so does this entry.
Nvidia said the scenarios were already known and that its memory protections block them. Eclypsium said those protections only help if they are configured and used, and that it has been working with Nvidia on them.
Why it matters
Most AI-safety debate is about model behaviour and lab practices. This research points to a different layer: the integrity of the servers that run the models, where a quiet change can alter answers to many users. The evidence so far is a vendor demonstration reported by one outlet, with two outside experts vouching for its plausibility; no paper or advisory is public yet.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 press
Press
Changes
- Filed (CNN read in full)