Microsoft's Agent Framework PyPI package compromised to steal credentials
MultiNodeTrainingValidation detects instances that pass hardware diagnostics but produce invalid training results.
To get an alert, create a CloudWatch metric filter for HANG_DETECTED events and an alarm that notifies an Amazon Simple Notification Service (Amazon SNS) topic:
aws logs put-metric-filter \ --log-group-name /aws/sagemaker/Clusters/<cluster-name>/<cluster-id> \ --filter-name hung-jobs \ --filter-pattern '{ $.event_type = "HANG_DETECTED" }' \ --metric-transformations \ metricName=HungJobsDetected,metricNamespace=HyperPod/HungJobs,metricValue=1
Turn off hung job detection To turn off detection for a job that would otherwise inherit HYPERPOD_HANG_RULES , add --no-hang-detection to srun , or to sbatch to cover every step in the job.
It also wrote the failure to CloudWatch with the pods and nodes involved, so you don't need to look them up from the job's pod list:
JSON Duality Views let applications read and write the same relational tables as JSON documents.
When you do need a model, right-size it Choose the cheapest and fastest model that meets your quality floor.
It also keeps a small in-process cache, so payloads already read or written during the current invocation do not require a second S3 round trip.
To support this, we introduced a dedicated persisted-state schema that keeps only lightweight fields and artifact references.
We implemented three strategies to reduce the token costs, latency, and memory overhead caused by passing large JSON payloads through the model context.
We also added basic statistics and cache data to state for rich queries.
We defined a custom agent state that holds the data needed by our tools: optimization request JSON, optimization response JSON, and validation report JSON.
Fixed a rare display fault that could freeze the screen or end the session - Fixed cloud sessions breaking when a program left thousands of connections open to the session's network proxy - Fixed a turn hanging when a mod's $.model.complete or $.model.classify call met a usage limit: the call now gets the API's error at once; mods should handle that error - Fixed stalls on a compaction reply with thousands of unclosed tags and on some very long WebFetch URLs; WebFetch now refuses a URL over 2,000 characters before any permission prompt - Fixed a message with thousands of @ mentions stalling before Claude replies; only the first 100 mentions are now looked up as files - Fixed the prompt freezing after rejecting a write to a very large file; a rejected write now shows a short preview - Fixed a crash when a burst of keys reached the search box in /plugin - Fixed a false crash report and exit code when a suspended session resumes, or at exit, where a Linux security policy blocks a process from signalling itself - Fixed remote MCP servers staying disconnected after a longer outage such as a VPN drop: most servers on other machines reached by hostname over https are now retried in the background for about a day - Fixed Edit refusing with "File has not been read yet" after a compaction on unchanged files outside the project that Claude had already read in full - Fixed Grep repeating or skipping results when paging with head_limit and offset; content and count results are now sorted
command running on an extra Enter after the usage limit menu - Fixed some --disallowedTools rules with parentheses in their content - Fixed auto mode ignoring some wildcard PowerShell allow rules that name a script with .\, such as PowerShell(.\scripts\build.ps1 *) - Fixed a running Code Mode script in auto mode getting permission prompts and denials once Claude started a second script that auto mode also allowed - Changed some pre-started background sessions to be discarded after an account change - Changed a workflow paused at a usage limit in the terminal to stay paused after a login, until the earlier login returns or you type a prompt - Fixed a nested claude with its own token being refused by forceLoginOrgUUID: commands that don't get the session's token no longer get its CLAUDE_CODE_OAUTH_SCOPES; set it on any that need it - Fixed headless sessions losing stored messages on resume when a message was sent again with the same UUID after the process was killed - Fixed cases where a session resumed after an empty reply or a killed command lost an answer, or brought back a rewound prompt, after another crash or a move to the background - Fixed a remote session forgetting messages sent after Stop was pressed during a restart, if Claude Code then crashed - Fixed tasks created after /clear being lost, and cleared tasks coming back, when a cloud or self-hosted runner session with no feature flags loaded resumed later - Fixed a compaction right after a reply cut at the output to
Fixed stopping the server in worktree mode deleting a running session's git-ignored files, such as .env - Fixed self-hosted runners with --use-anthropic-git-proxy deleting a non-empty directory named by GIT_CONFIG_GLOBAL at startup: the runner now keeps it, ignores the variable and warns - Fixed claude mcp add and claude mcp remove with project scope deleting other top-level keys from .mcp.json, such as $schema or another tool's servers - Fixed /export overwriting an existing file without asking: the dialog now asks, and /export <name> keeps the file and says to run /export - Fixed a command you had allowed still running after you pressed Stop, when the approval arrived after a restart, in cloud and self-hosted runner sessions with no feature flags loaded - Fixed Enter closing /teleport's load error screen - Fixed a click followed by Enter in /plugin installing or disabling a plugin - Fixed a half-typed !
Microsoft's Agent Framework PyPI package compromised to steal credentials