Artificial intelligence is moving into increasingly sensitive operational environments, but the systems used to evaluate AI safety are still struggling to scale alongside deployment. As enterprises adopt large language models across manufacturing, IT services, and workflow automation, Korean startup Selectstar (Datumo) is attempting to address a growing bottleneck: how to test AI systems continuously and efficiently before they reach real users.
Selectstar (Datumo)’s STAR-Teaming Research Reaches ACL 2026 Findings
South Korean AI reliability startup Selectstar announced that its automated AI red-teaming framework, “STAR-Teaming,” was accepted as part of ACL 2026 Findings, one of the leading academic venues in natural language processing research.
The paper, titled “STAR-Teaming: A Strategy-Response Multiplex Network Approach to Automated LLM Red Teaming,” focuses on automated methods for identifying vulnerabilities in large language models (LLMs). According to the paper’s arXiv record, the framework was tested across 17 LLMs, including models such as Claude, ChatGPT, Gemma, Llama, and Qwen.
Red teaming refers to adversarial testing designed to intentionally provoke unsafe, harmful, or misleading model responses before AI systems are deployed publicly. The process has become increasingly important as generative AI systems move beyond experimental chatbots into customer-facing services and operational enterprise environments.
Why Automated AI Red Teaming Is Gaining Attention
The broader AI industry has been facing a practical scaling problem around AI safety evaluation.
Manual red teaming can require significant time, specialized expertise, and repeated testing cycles. At the same time, enterprises are deploying AI systems more rapidly across products, internal workflows, and automated services.
Selectstar’s research focuses on automating part of that evaluation process.
According to the paper, STAR-Teaming achieved an average attack success rate (ASR) of 74.5% on the HarmBench benchmark. The company said this represented a 13.5 percentage-point improvement over AutoDAN-Turbo, which recorded 61.0% in the same comparison setting.
Rather than repeatedly selecting previously successful attack examples, STAR-Teaming models relationships between attack strategies and model responses to identify more effective adversarial approaches probabilistically.
The research also aims to reduce the number of testing attempts required during red teaming, an issue that has become increasingly relevant as enterprises attempt to evaluate AI systems at larger scale.
Datumo Platform Connects Research With Enterprise AI Deployment
The significance of the announcement extends beyond academic publication.
Selectstar said the technology has already been integrated into its AI reliability evaluation solution, Datumo Platform, where it is being used in automated red-teaming workflows for enterprise AI validation.
According to the company, the technology is currently being applied across sectors including electronics manufacturing, home appliances, system integration (SI), and IT services, as well as government-led sovereign AI foundation model initiatives in Korea.
This reflects a broader industry transition where AI safety evaluation is becoming part of operational deployment infrastructure rather than remaining a research-only function.
In a follow-up response to ngopihangat, Selectstar said automated red teaming is currently used most often before AI services launch in order to identify safety and quality risks prior to deployment.
The company also noted that enterprises increasingly continue safety evaluation after launch when AI systems are updated, renewed, or expanded.
“Companies use it to identify potential safety risks and quality issues before their AI services are released to users,”
Selectstar told ngopihangat.
“It is also being used after launch, particularly when companies renew, update, or further enhance their AI services.”
Enterprise AI Deployment Is Increasing Pressure on Safety Evaluation
The demand for automated AI evaluation is growing alongside enterprise adoption of generative AI systems.
Industry groups such as OWASP and major AI companies including Google and Anthropic have increasingly emphasized red teaming, adversarial evaluation, and safety testing as core components of AI deployment practices.
At the same time, enterprises face operational challenges in maintaining those evaluation systems internally.
In its correspondence with ngopihangat, Selectstar said manpower and cost remain major bottlenecks for enterprise AI safety evaluation. The company also pointed to the speed at which red-teaming methodologies continue evolving.
“As more effective approaches become available, existing methods can quickly become outdated, which means companies need to continuously develop and adopt improved evaluation methodologies,”
the company said.
Selectstar added that the pool of specialists with advanced AI red-teaming expertise remains relatively limited, creating additional hiring and operational pressures for enterprises attempting to expand AI deployment safely.
Korea’s AI Reliability Startups Are Expanding Beyond Data Annotation
The development also highlights how parts of Korea’s AI startup ecosystem are evolving beyond traditional data labeling and annotation services.
Selectstar originally built its business around AI data infrastructure and validation workflows. More recently, the company has expanded into AI reliability testing, LLM evaluation, and automated red teaming through its Datumo Platform.
The shift reflects a wider change occurring across the global AI ecosystem. As enterprises move AI systems into production environments, the challenge is no longer limited to training models. Increasingly, companies also need scalable ways to evaluate reliability, safety, and operational behavior throughout deployment cycles.
For Korean AI startups, this creates opportunities in infrastructure layers that sit around AI deployment rather than competing directly in frontier model development itself.
AI Safety Testing Is Becoming Part of Enterprise Infrastructure
Finally, this ACL 2026 Findings acceptance gives Selectstar academic visibility, but the company’s larger positioning appears tied to the operational deployment needs inside enterprise AI systems.
Automated red teaming is increasingly being treated less as a one-time compliance exercise and more as an ongoing evaluation layer attached to AI deployment cycles.
As enterprises continue integrating LLMs into products and operational workflows, demand for scalable safety testing frameworks may continue expanding alongside AI adoption itself.
And Selectstar’s latest research suggests that Korean AI startups are beginning to position themselves inside that emerging infrastructure layer.
Key Takeaways
- Selectstar (Datumo)’s STAR-Teaming framework was accepted as part of ACL 2026 Findings, a leading academic venue in natural language processing research.
- The company reported 74.5% attack success rate (ASR) on HarmBench, compared with 61.0% for AutoDAN-Turbo in the paper’s benchmark comparison.
- STAR-Teaming has been integrated into Datumo Platform and is being applied in enterprise AI validation workflows across manufacturing, SI, and IT services.
- Selectstar told ngopihangat that automated red teaming is used both before AI deployment and after service updates, reflecting growing demand for continuous AI evaluation.
- Selectstar (Datumo)’s story reflects a broader industry trend where AI safety testing is becoming operational infrastructure as enterprises deploy generative AI systems at larger scale.
🤝 Looking to connect with verified Korean companies building globally?
Explore curated company profiles and request direct introductions through beSUCCESS Connect.
– Stay Ahead in Korea’s Startup Scene –
Get real-time insights, funding updates, and policy shifts shaping Korea’s innovation ecosystem.
➡️ Follow ngopihangat on LinkedIn, X (Twitter), Threads, Bluesky, Telegram, Facebook, and WhatsApp Channel.
PakarPBN
A Private Blog Network (PBN) is a collection of websites that are controlled by a single individual or organization and used primarily to build backlinks to a “money site” in order to influence its ranking in search engines such as Google. The core idea behind a PBN is based on the importance of backlinks in Google’s ranking algorithm. Since Google views backlinks as signals of authority and trust, some website owners attempt to artificially create these signals through a controlled network of sites.
In a typical PBN setup, the owner acquires expired or aged domains that already have existing authority, backlinks, and history. These domains are rebuilt with new content and hosted separately, often using different IP addresses, hosting providers, themes, and ownership details to make them appear unrelated. Within the content published on these sites, links are strategically placed that point to the main website the owner wants to rank higher. By doing this, the owner attempts to pass link equity (also known as “link juice”) from the PBN sites to the target website.
The purpose of a PBN is to give the impression that the target website is naturally earning links from multiple independent sources. If done effectively, this can temporarily improve keyword rankings, increase organic visibility, and drive more traffic from search results.
