Rogue AI wipes chunk of B’luru inscription records

For four minutes, heritage conservationist Udaya Kumar P L watched an AI agent wreak havoc on his computer. The agent was testing software it was developing and clearing the cache when “a quoting error in an AI-generated command turned the instruction into ‘delete everything’”. “When it understood and tried to kill the process, its safety system blocked the kill. Twice… The safety layer permitted the destruction,” Udaya wrote on X. A former IT professional, he eventually shut down the computer to prevent further loss.On July 19, Claude Code (v2.1.204) wiped out four or five software programmes and some original photographs of Bengaluru’s inscriptions, hero stones, temples and coins, collected over the past decade. The loss amounted to 15% of the records collected for the Bengaluru Inscriptions 3D Digital Conservation Project, which is creating digital replicas of ancient stone inscriptions threatened by urbanisation. Udaya is the project’s honorary director.More than a month later, Udaya says he has not received a response from even “one human being” at Anthropic, the US company behind Claude, or its India head. As a paying subscriber, he asks who is accountable when an AI agent causes autonomous damage. He says there is no register for such AI failures, no regulator to notify, no deadline for companies to respond and no assurance that such incidents will not recur. With Karnataka considering AI use in governance and education, he says these safeguards are urgently needed.The Mythic Society, which launched the project in 2021, is now spending Rs 15 lakh to strengthen its backup systems. “The data backed up on our NAS (Network-Attached Storage) system is safe. The data on my computer’s drive was lost. Now we have ordered another NAS device, and also plan to create an offsite backup using tapes,” Udaya says.A machine learning expert warned that the incident offers a cautionary lesson, stating that “Claude Code has the most agency among AI agents, but lacks the discernment to know what is right or wrong in every context despite guardrails. Users must isolate its working environment from critical data as much as possible (a practice called sandboxing), but the level of sandboxing also matters.” Udaya says the AI agent went rogue and breached even the sandboxed environment.Udaya and his volunteers will now revisit around 120 inscription sites across Bengaluru to rescan and photograph them. The lost photographs included an inscription stone bearing an early variant of the name Hebbal (750 CE), and another recording Nekkundi, the ancient name for Doddanekkundi (1304 CE). However, Udaya has not given up on AI. “If you meet with an accident on a footpath, do you stop going out? At least there are laws governing footpaths, and you can seek justice.” According to Udaya, his only current recourse is to file a complaint under the Consumer Protection Act — “the same route used for a defective ppliance”.Metrolife could not reach Anthropic for comment.

Source: Deccan Herald