Claude Code soft_deny policy hits human review gap

// 1h agoNEWS

Claude Code soft_deny policy hits human review gap

Hedgineer's enterprise rollout of Claude Code reveals that natural language 'soft_deny' rules, while doubling automated rejections, fail to catch many risky bash commands that developers still manually veto. The findings highlight a persistent gap between AI intent classification and human risk assessment in autonomous coding environments.

// ANALYSIS

Automation is only as good as its telemetry, and currently, Claude's 'soft_deny' is a blunt instrument that misses subtle context.

–Soft_deny rules are bypassable by explicit user intent, making them "negotiable" guardrails rather than hard blocks
–Classifier-driven rejections jumped 123% post-policy, yet Bash remains the top tool rejected by humans in the loop
–Current OTEL spans don't distinguish between hard, soft, and permission denials, making it impossible to surgically tune rules
–The "trap" of omitting "$defaults" in config can inadvertently allow dangerous operations like force pushes
–Enterprise safety relies on identifying "bad vibes" in telemetry and encoding them back into natural language policy

// TAGS

claude-codeai-codingagentsafetyobservabilitymcpdevtool

DISCOVERED

1h ago

2026-05-30

PUBLISHED

3h ago

2026-05-30

RELEVANCE

8/ 10

AUTHOR

dani_avila7

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE36m ago

Claude Code defaults to Opus 4.8

Claude Code v2.1.154 promotes Opus 4.8 to the default high-effort model, adds dynamic workflows that can orchestrate work across dozens to hundreds of background agents, and improves fast mode economics and speed on Opus 4.8. The release also refines cleanup flows with a lighter `/simplify` path, renames effort labels for clarity, and tightens several CLI and agent workflows for heavier terminal-based coding sessions.

TUTORIAL45m ago

LLMWhisperer powers complex-document RAG pipelines

The video shows LLMWhisperer as Unstract’s layout-aware text extraction layer for PDFs, images, and scanned documents. That preprocessing step turns messy files into LLM-ready input for downstream extraction and RAG workflows.

TUTORIAL45m ago

Unstract tutorial covers local setup

This YouTube walkthrough shows how to self-host Unstract, the open-source document extraction platform, with Docker and local model support. It positions the tool as a practical fit for offline and private RAG-style workflows that turn PDFs and other files into structured outputs.

Claude Code soft_deny policy hits human review gap