Researchers Discovered BadEngram AI Supply-Chain Attack
A newly identified threat inserts persistent, dormant backdoors into open-weight AI models through memory modifications.
Updated on Oct. 13, 2026 in Artificial Intelligence
Live Poll
Do you trust the security of open-weight AI models for your business or personal use?
Researchers have uncovered a supply-chain attack known as BadEngram that targets gated memory architectures in open-weight artificial intelligence models. The method allows attackers to implant a persistent, dormant backdoor that activates only when triggered by a specific, arbitrary string.
Why it matters
This vulnerability is significant because gated memory components allow AI models to retrieve information from dedicated storage, creating a potential security loophole that is difficult to detect through standard model testing.
The BadEngram attack exploits gated memory architectures by modifying only memory-specific parameters, leaving the core computation of the model untouched. This surgical approach ensures the backdoor remains dormant until triggered by a specific short string.
The players
Pillar Security
This is a cybersecurity firm that specializes in identifying vulnerabilities in emerging technology infrastructures.
Fujitsu
This is a multinational information and communications technology equipment and services corporation.
The details
By specifically altering memory components, the attack circumvents conventional detection methods that monitor primary model weights. Pillar Security and Fujitsu researchers demonstrated how this persistent backdoor bypasses standard security checks during typical model operation.
Timeline
Pillar Security and Fujitsu announced the BadEngram discovery on October 13, 2026.
The Tech Race
The discovery of BadEngram highlights the emerging arms race between AI developers and threat actors as architectures move beyond simple neural networks. This vulnerability signals a departure from traditional model poisoning toward sophisticated memory-level manipulation.
Users and developers of open-weight models may face increased risks of supply-chain contamination that standard security tools cannot currently detect. This necessitates a shift toward more rigorous, specialized auditing of model memory components before deployment.
The takeaway
Developers should prioritize testing the integrity of memory-specific parameters when incorporating gated architecture models. This discovery serves as a reminder that securing the data-retrieval process is as critical as securing the model weights themselves.
Further reading
Learn more about the latest developments in Artificial Intelligence.
Source note: This article includes information reported by The Manila times.
Live Poll
Do you trust the security of open-weight AI models for your business or personal use?
