AINEWS
Hours after OpenAI tightened agent security and Anthropic loosened its own, developers pressed the same vendors — and Cursor — to flip their AI coding assistant defaults to secure by default.
Hours after OpenAI moved to tighten agent security and Anthropic loosened its own, a coordinated group of developers moved in the opposite direction: a public petition and a public dataset directed at the same vendors and their peers, asking them to flip the default settings on their AI coding assistants to safe-by-default.
The Register on Friday framed the story around a preprint called "Impossible to hide secret ..." (arXiv:2607.26390), which systematically catalogued security and privacy reports across popular LLM-native IDEs, or LIDEs. The paper's authors scraped 1.1 million posts from 29 LIDE-related subreddits, narrowed them to 446 substantive discussions, and coded more than 6,000 comments on Cursor, Copilot, Codex and other commercial coding assistants.
The authors built a taxonomy of reported issues covering nearly every popular LIDE and released the data and labels in a public replication package on GitHub. The taxonomy groups the complaints: telemetry settings that ship permissive, code-context sharing that streams more than the user expected, secret and PII handling that does not warn before transmission, and permission scopes that are too wide to be safe. The finding, in plain language, is that the default configuration of these tools is doing too much unstated work. When telemetry, context sharing, and permission scopes are all permissive out of the box, "user-friendly" and "user-exposed" describe the same thing.
The developer petition The Register reported carries the same mechanism into a concrete demand. The signatories are not asking the labs to add a privacy panel. They are asking Anthropic, OpenAI, Cursor and their peers to flip the default: safe code generation, scoped permissions, and explicit secret and PII handling should be the out-of-the-box posture, with the permissive option moved behind an explicit opt-in. The petition frames it as a competitive and trust question, not a compliance one. The labs that ship secure-by-default, the argument goes, win the enterprise and security-engineering buyers; the ones that ship permissive-by-default inherit the risk.
Defaults matter because the majority of users never change them. A vendor that sets a permissive default is picking a threat model on the user's behalf. The Register coverage and the arXiv taxonomy both lean on that mechanism rather than on any single incident, because the data is from public discussion. 446 flagged posts across 29 communities is large enough to read as a signal, not large enough to settle any specific exposure claim.
The study is a preprint, not peer-reviewed. Its taxonomy is the authors' coding of community-reported problems, and Reddit selection and survivorship bias apply. The Register excerpt available to this piece is short and dominated by site navigation; the full article's named developer demands should be checked before quoting. The picture is still legible: a developer constituency is on the record, naming specific vendors, asking for a specific posture change, and pointing to a public dataset that catalogs the problems.
The next move is product, not press. A roadmap line about "safer defaults" is not the same as flipping a telemetry switch, narrowing a permission scope, or moving secret handling behind a confirmation. The arXiv authors made their data and taxonomy public. The petition made the ask public. The remaining question is whether the defaults page on a Cursor or Codex install actually changes.