Naver Cloud Corp. said Friday it tightened default guardrails on Gauss enterprise large language models after a domestic phishing simulation showed white-collar workers could trick internal copilots into repeating confidential contract clauses.
What failed in the drill
The Korea Internet & Security Agency ran a two-week exercise with 14 listed companies, asking employees to interact with sandboxed assistants trained on anonymized HR and procurement documents. In 11 percent of attempts, participants posing as executives elicited language that matched restricted templates closely enough that compliance officers flagged it as a data leak.
Naver Cloud, which hosts several participants’ models, said it will enable stricter prompt filtering and session logging by default for new Gauss deployments starting Nov. 1. Existing customers can opt in immediately through the console.
Technical changes
The update adds retrieval-time redaction for national ID patterns, bank account numbers, and project codenames stored in vector databases. Naver also reduced maximum context windows for email-plugin integrations, limiting how much thread history a model can see per query.
Regulatory context
Korea’s Personal Information Protection Commission reminded firms that treating generative AI like a search engine does not relieve them of breach-notification duties. The commission is drafting guidance on log retention for enterprise models, expected before year-end.
Competitors including KT Cloud and NHN Cloud said they are reviewing similar defaults. Industry groups argued that overly aggressive filtering could blunt productivity, setting up a debate similar to email DLP rollouts a decade ago.
Enterprise adoption
Despite the drill, Naver reported rising Gauss subscriptions from insurers and municipal governments automating citizen FAQs. Chief product officer Lee Hye-soo said customers demanded clearer toggles between “creative” and “compliance” modes rather than a single global setting.
For Korean offices rolling out copilots during APEC security lockdowns, Friday’s patch is a reminder that model safety is as much about social engineering as it is about GPUs.
Customer rollout
Several Gauss early adopters in the insurance sector said they will run parallel models for claims adjusters—one filtered, one experimental—until year-end audits finish. Naver Cloud product managers acknowledged the dual-track approach increases GPU costs but prevents business units from bypassing safety teams through shadow IT accounts.
Training materials updated Friday include sample prompts that attackers used in the KISA drill, redacted to remove company names. Security vendors selling prompt-firewall appliances reported inbound leads from Korean banks that had previously delayed purchases.
Longer-term roadmap
Naver plans to certify Gauss deployments under ISO 42001 AI management standards by mid-2027, a timeline competitive with global hyperscalers. Until then, chief information security officers said they will treat copilots like privileged admin accounts, requiring multi-factor authentication for every session.
Sector uptake
Municipal governments in Gyeonggi Province said they will delay citizen chatbot rollouts until Gauss guardrails pass third-party penetration tests. Naver offered discounted audit packages to public-sector clients.
Education ministry officials asked edtech vendors using Gauss to disable student-facing plugins until winter break, citing exam-season cheating concerns separate from the phishing drill.
Naver said it will publish monthly aggregate statistics on blocked prompts to build trust with regulators without exposing customer data.
Large law firms in Seoul updated client memos Friday advising companies to log every Gauss query during litigation holds, treating model outputs as potentially discoverable documents.
Rival products
Open-source model hosts in Korea rushed blog posts explaining their own safety layers, arguing that enterprises should compare audit trails rather than defaulting to brand names alone.