Post by Udaya Kumar PL: On 19 July, an AI agent deleted more than five years of heritage documentation from my computer. I have spent 22 days waiting for one human being at @AnthropicAI to say a single word to me. I was using Claude Code (v2.1.204), Anthropic's coding agent, to clear a cache. It wrote the command itself. A quoting error turned that command into delete-everything, and it ran in the background. It kept deleting for about four minutes while the AI was still trying to work out why files were disappearing. It did not realise its own command was the cause. When it finally understood and tried to kill the process, its own safety system blocked the kill. Twice. The deletion stopped only when I shut the machine down by hand. Read that again. The safety layer permitted the destruction, then stood between me and the off switch. My drives are SSDs with TRIM. The blocks were physically erased within minutes. Recovery tools show empty space. What was lost is not really mine. It belongs to @MythicSociety, documenting India's epigraphic and numismatic heritage - photographs of inscriptions, temples, herostones, coins, catalogued over more than five years, largely by me and some volunteers. Some of it was the only photographic record we held. I have had to tell those volunteers their work is gone and must be done again. That conversation is what I would like Anthropic to picture. I am a paying subscriber. I reported it the same night with a full technical write-up. An automated acknowledgement arrived two minutes later. In 22 days, that remains the only response I have ever received.

A nudge on 21 July. A formal follow-up to usersafety@ on 28 July. Messages on Discord. An in-product bug report with the entire session transcript attached. A public GitHub issue, #82165, on 29 July. Nothing. Here is what should concern everyone, far more than my loss: When an autonomous AI agent damages you, nothing obliges the company to respond at all. There is no regulator to notify. No incident register. No timeline anyone must meet. The entire cost sits with the user. In India, my realistic recourse is a complaint under the Consumer Protection Act - the same route as a defective appliance. That is the whole legal apparatus available when software that acts on its own destroys irreplaceable work. These tools are now sold to run unattended on people's machines. If that is the product, four things should be law rather than goodwill: * a mandatory, time-bound response to reported harm * a reportable-incident register for autonomous agents, as aviation and medicine have * clear liability for actions an agent takes on its own initiative * safety controls that can never prevent a user from stopping the software Until then, every user is a test subject bearing the full cost of the experiment. To be clear about what I asked for, in writing, on the night it happened: reimbursement of the costs I incurred - data-recovery software, and the substantial time and effort of rebuilding the systems and data that were destroyed - along with consideration for work that is permanently lost. And an account of what has changed, so this cannot happen to the next person. I have not had a reply to any of it. Not a refusal. Not a question. Nothing. @AnthropicAI @ighose - I am still waiting. Technical detail, exact command, process output and the blocked kill attempts: