What AI Update Did You Miss Today?

Claude Code Closes Auto-Approve and Prompt-Check Gaps in BASH_ARGV0, Interrupted Hooks and Git Bash rm

Claude Code 2.1.296 patches a batch of cases where the agent could act outside what a user authorized. Bash commands that assign `BASH_ARGV0` and then use it are no longer auto-approved, an Esc or interrupt during a prompt-submit hook can no longer let an unchecked prompt through, and managed hooks that deny a tool call now end the turn properly. The release also restores the ask prompt for `rm -rf` on a Windows user folder in Git Bash under bypass permissions mode, and makes cloud sessions run auto mode's checks on Claude in Chrome actions even before feature flags load.

Claude Code

Claude Code Edit Tools Now Refuse to Overwrite Non-UTF-8 Files

Claude Code 2.1.296 changes how the Edit and NotebookEdit tools treat files that are not valid UTF-8, such as Windows-1252, Shift-JIS and GBK files. Previously an edit could replace every non-ASCII character in such a file, silently damaging text far from the lines being changed. Edits to these files are now refused instead.

Claude Code

Claude Code Cost Figures Now Reflect Cheaper Sonnet 5.5 Cache Reads

Claude Code 2.1.296 updates `/cost`, the status line, `--max-budget-usd` and the SDK's cost figures to price Claude Sonnet 5.5 cache reads at $0.10 per million tokens, down from $0.20. Because Claude Code leans heavily on prompt caching in long sessions, the lower rate shows up directly in reported spend and in how far a budget cap stretches.

Claude Code

Claude Code Hooks Can Now Fail Closed With onFailure

Claude Code 2.1.295 adds an `onFailure: "block"` option for command and HTTP hooks. When set, a hook that cannot start, times out, or exits with an unexpected code blocks the action instead of letting it through. This closes a long-standing gap where a broken guard hook silently stopped guarding anything.

Claude Code

Claude Code Fixes Prompt and Agent Hooks That Failed to Block Actions

Claude Code 2.1.294 fixes a gap where `prompt` and `agent` hooks written as instructions, such as "Block commands that...", could allow the very actions they were meant to block. The release also changes how Stop and SubagentStop `prompt` hooks written as instructions, such as "Carry on if the build is broken", are judged. Claude is now less likely to stop early.

Claude Code

Claude Code Fixes Deny Rule and Tool Restriction Gaps

Claude Code 2.1.295 closes several gaps where a user-configured guardrail did not block what it should. A Chrome site deny rule written with port 80 was ignored on plain http pages, `--tools` and `--restricted` missed late-registering built-in tools, and some mod guard hooks could be bypassed or shown truncated input. A Bash permission check for for-loops over globs was also tightened.

Claude Code

Claude Code Adds Claude Haiku 5.5 as the Default Haiku Model

Claude Code 2.1.293 adds Claude Haiku 5.5 (`claude-haiku-5-5`), which is now the default Haiku model on the Anthropic API. The model offers a 1M-token context window and is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens. Prompts over 100K tokens cost $0.50 and $2.50. That is a large price cut from Haiku 4.5, which cost $1 and $5 per million tokens.

Claude Code

Claude Code: Agent Tool Gets an Effort Parameter for Sub-Agents

Claude Code 2.1.292 added an `effort` parameter to the Agent tool, so Claude can run a sub-agent at the effort level a user asks for. Delegated work no longer has to inherit the main session's effort setting.

Claude Code

Claude Code: Permission and Sandbox Gaps Closed Across Auto Mode, UNC Reads and Skills

Claude Code 2.1.292 closes a batch of gaps where the agent could act beyond what a user authorized, including network (UNC) file reads that skipped the permission prompt and subagents entering auto mode when it was unavailable. Several sandbox read-deny and link-swap issues were fixed alongside them.

Claude Code

Claude Code Closes Permission and Read-Block Bypasses

Claude Code 2.1.290 patches a batch of cases where deny rules, read blocks and auto-approval could be sidestepped. Affected paths include variable-prefixed `declare` and `export` commands, pasted image paths, `rg` and `git grep` wildcards, zsh variable parsing, and symlinks swapped mid-read. It also stops project settings from enabling Claude in Chrome and asks before running `pyright` or some `ps` forms.

Claude Code

Claude Tag Adds !fast Mode for Slack Threads

Claude Tag in Slack gains a fast mode. Mentioning Claude with `!fast` switches a thread to fast mode, moving it to Opus if needed, and `!fast off` switches back. Replies show "(fast)" while it is on.

Claude Code

Claude Code Scheduled Tasks Survive Compaction and Resume

Claude Code 2.1.290 fixes `/loop` and reminder tasks silently disappearing after conversation compaction or a background hand-off. Recurring tasks no longer fire an extra run on every resume, respawn or fork, and background sessions waiting on a scheduled wakeup are no longer shut down by updates or low memory.

Claude Code

Claude Code WebFetch Now Reads Past 100,000 Characters

Claude Code 2.1.290 fixes WebFetch silently dropping page text beyond 100,000 characters. The tool now reports how much was left unread and accepts an `offset` parameter so Claude can read on.

Claude Code

Claude Code WebSearch Budget Refills 100 Calls Per Hour

Claude Code 2.1.290 changes the interactive session's WebSearch limit. Instead of ending after 200 calls, the budget now refills over time at 100 calls per hour, adjustable with an environment variable.

Claude Code

Claude Code /ultrareview Upload Fixes for Uncommitted Work

Claude Code 2.1.290 fixes several ways `/ultrareview` could drop, mis-filter or refuse uncommitted changes, including on Windows, with `core.safecrlf=true`, and with git 2.54+ layouts. Upload errors now name the failing git step and what to try.

Claude Code