Urgent.News

What's breaking now, across thousands of outlets.

AI

The failures that don't fail loudly

I spent a week building an agent that upgrades dependencies and repairs what the upgrade breaks. Dependabot opens the PR and walks away; this one stays until the tests are green. The interesting part wasn't the repairs. It was that almost every serious bug I hit — in my code, in the harness, in my own agent — announced success while being wrong. The premise Detection is solved. Dependabot, npm…

In a week-long project, a reporter built an agent designed to upgrade dependencies and repair any broken code resulting from the upgrades. Unlike other automated tools like Dependabot, npm audit, and OSV, this agent didn't simply detect issues but actively attempted to fix them. However, the reporter discovered a series of significant bugs within the system that didn't seem to get caught by the usual detection methods.

The first issue was that the scanner was not detecting vulnerabilities in the lockfile, which contains information about resolved dependencies and their transitive dependencies. The reporter learned that to accurately detect vulnerabilities, the scanner must scan the entire lockfile, not just the package.json file.

The second problem was that the agent itself had a bug, which led it to report 124 vulnerable packages instead of the actual 3. The agent had nested the version inside the package object instead of placing it alongside the package name, causing the API to return every advisory for each package. The reporter fixed this issue with three separate subagents, each responsible for fixing one advisory. However, one of the fixes ended up getting silently overridden by another fix, resulting in the first fix being lost.

Finally, the reporter discovered that there were several ways the scanner could under-report vulnerabilities, including using aliases, pagination, link entries, truncated responses, and the gate working but its display not functioning properly. The reporter implemented changes to address these issues, such as refusing to continue if the response didn't match the query, and explicitly instructing the agent that being blocked was a correct outcome in the face of irreversible actions.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Three layers of automated fact-checking for an LLM newsroom (and the bugs that forced each one)

Our site, presentofai.com , publishes AI industry analysis daily with no human in the writing loop: agents ingest news and company filings into an event timeline, score them, and synthesize digests…

  • First layer, article-level critic checks drafts against source events for errors
  • Second layer, search-verified claim checking extracts and verifies crucial claims
  • Third layer, data auditing re-reads timelines for recurring failure classes

The agent wrote the migration. Are you really going to Approve it?

Tuesday afternoon, that Approve button Tuesday, 3 p.m. A PR notification lands in Slack. Junior writes in the channel: "Cursor generated this migration for me — can you take a look?" You open the…

  • Agent autonomously generated 80-line migration with ALTER TABLE statements
  • Migration issues: missing default value, table lock risk, inaccurate down migration
  • PR description insufficient; proper scrutiny needed before approval

More from Sunday 30 August →