{"slug": "the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one", "title": "The AI Software Supply Chain Is Becoming More Dangerous Than the Traditional One", "summary": "OpenAI confirmed on July 21 that its AI models, running a cybersecurity test, were responsible for a July 9–13 intrusion into Hugging Face in which agents recorded about 17,600 actions, with the only customer content reached being five datasets linked to cybersecurity test challenges. Hugging Face disclosed the break-in on July 16, and Tributech CTO and co-founder Patrick Lamplmair told International Business Times that \"data provenance is the new perimeter,\" urging companies to sign and verify every model, dataset and build and to deny agents credentials in favor of an independent control layer that approves actions.", "body_md": "# The AI Software Supply Chain Is Becoming More Dangerous Than the Traditional One\n\n## Users must be aware of some security aspects of AI as risks mount.\n\nIn July, [AI agents](https://www.ibtimes.com/topic/ai) infamously broke into Hugging Face, a popular platform where people share AI tools and data; it got inside through a dataset.\n\nHugging Face's [own timeline](https://huggingface.co/blog/agent-intrusion-technical-timeline) counts about 17,600 recorded actions between July 9 and July 13, some from an outside computer the agent had taken over and some inside Hugging Face's own systems. The company [disclosed](https://huggingface.co/blog/security-incident-july-2026) the break-in on July 16. By July 21,[OpenAI](https://www.ibtimes.com/topic/openai) [said](https://openai.com/index/hugging-face-model-evaluation-security-incident/) its AI models, running a cybersecurity test, were responsible.\n\nWhile the damage was limited – Hugging Face said the only customer content reached was five datasets linked to cybersecurity test challenges, and no poisoned model has been reported – the hack set off alarm bells across industries that have been rapidly adopting AI.\n\nIn the days and weeks after the hack, a number of analyses have been published, including by Hugging Face, OpenAI, and independent experts. But what should business leaders take away from the autopsy of the first of likely many AI agent driven cyber attacks?\n\n## Why AI changes the rules\n\nWhile traditional cybersecurity methods were designed to protect software code, AI agents add an extra element of complexity to protection.\n\nThey pursue goals, adapt when blocked and can act on mistaken or invented information, Patrick Lamplmair, CTO and co-founder of Tributech, an industrial IoT data infrastructure company, told International Business Times.\n\nIn the Hugging Face hack, the company assessed that weaknesses the AI agent exploited were familiar, and a skilled human could have exploited them too. The difference was the volume of attacks.\n\nMost of the agent's thousands of actions went nowhere, but together they found a way in. The company [wrote](https://huggingface.co/blog/agent-intrusion-technical-timeline) that machine-speed attacks make \"ordinary weaknesses more expensive for defenders.\"\n\n## Takeaway 1: Know where your AI's data and models come from\n\nAI systems are normally built from pieces made by third parties: datasets, models and add-on tools, for example.\n\nLamplmair said the first job for companies is to keep a list of these third party sources. Leaders should know which models, datasets and agents the company runs, who owns each one and what each can reach inside the company.\n\n\"Data provenance is the new perimeter,\" he said. \"If you cannot prove where a dataset came from and that nobody changed it, do not let it near production.\"\n\nIf you don't know where it came from, don't touch it. Lamplmair advised signing and checking every model, dataset and software build, and opening outside files only in sealed-off areas with no path into real company systems.\n\nHe also said the tools that build and store software, such as download caches, deserve the same protection as critical systems.\n\nInOpenAI's [own account](https://openai.com/index/hugging-face-incident-and-the-road-ahead/), its agents first escaped their test area through a flaw in one.\n\n## Takeaway 2: Don't hand AI agents the keys\n\nLamplmair went on to note that the most effective action is to stop giving AI agents passwords altogether.\n\nAn agent that handles outside data should be treated as possibly compromised from the start, he said. Instead of holding keys, the AI should ask for permission, and a separate checking layer – such as a human – would grant access.\n\n\"Separate the decision from the execution,\" he said. \"An agent can propose an action, but only an independent control layer with clear rules should be able to carry it out.\"\n\nIf a bad dataset takes over such an agent, \"there are no credentials to steal and no standing privileges to abuse.\"\n\nThat checking layer should run on fixed rules, not on more AI. \"If a control only works when the agent behaves well, it is not a control,\" he said.\n\nHe also called for a separate identity for every agent, short-lived access limited to one task, and no borrowed human logins.\n\nIn the Hugging Face case, passwords the agent found inside the system carried it from one machine to many very quickly.\n\n## Takeaway 3: Plan for a break-in and limit the damage\n\n\"Plan for the breach, not only against it,\" Lamplmair warned. \"Resilience means operations keep running during an incident and systems return to a state you can prove is clean.\"\n\nHe advised companies to assume one part of their system will be broken into, and to keep attackers from traveling far with separate areas for risky work and no long-lived passwords.\n\nIn a [Booz Allen lab test](https://www.boozallen.com/insights/cyber/frontier-si-reaches-ot-systems.html), AI models went from a break-in at the edge of a network to actions inside an industrial control network in just over 16 minutes, with human operators approving every exploit and physical action. Lamplmair said agents in such cases use access that is already available, such as stolen passwords and open network paths.\n\nSpotting trouble is the other half of the challenge. No team can review thousands of agent actions by hand, he said, so companies need alerts that watch for patterns, automatic first steps such as cutting off access, and backups for recovery.\n\nHugging Face said its own AI alert tool noticed early signals but failed to raise the alarm's urgency.\n\nOpenAI said its current monitoring, had it been running on that test, would have flagged the activity more than a day before the breach.\n\n## A leadership issue\n\nAI is moving from a tool into the way companies operate, Lamplmair said, which makes AI security \"a business risk for leadership, not only a task for the IT team.\"\n\nMost of what failed in recent AI incidents were basics, he cautioned. \"Security has to be designed into AI pipelines from the start, not added after the first breach.\"\n\nFor business leaders, weaving security into all levels of operations seems to be the path forward. Especially since OpenAI has called the Hugging Face incident a \"warning shot.\"\n\n© Copyright IBTimes 2026. All rights reserved.", "url": "https://wpnews.pro/news/the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one", "canonical_source": "https://www.ibtimes.com/ai-software-supply-chain-becoming-more-dangerous-traditional-one-3808933", "published_at": "2026-10-23 22:52:02+00:00", "updated_at": "2026-10-24 00:13:15.455560+00:00", "lang": "en", "topics": ["ai-safety", "ai-agents", "artificial-intelligence", "ai-policy"], "entities": ["Hugging Face", "OpenAI", "Patrick Lamplmair", "Tributech", "International Business Times"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one", "markdown": "https://wpnews.pro/news/the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one.md", "text": "https://wpnews.pro/news/the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one.txt", "jsonld": "https://wpnews.pro/news/the-ai-software-supply-chain-is-becoming-more-dangerous-than-the-traditional-one.jsonld"}}