Claude Code v2.1.325 Fixes Security and Permissions
Anthropic released Claude Code v2.1.325 on November 13, fixing a response-decoding security issue and changing how sandboxed commands receive permission. The update also repairs interrupted sessions, specialized agent behavior, and failures involving unusual Git repository layouts. Its central tension concerns everyday automation: stronger boundaries can change workflows that previously ran unattended.
The release targets the software surrounding Claude’s models, including command execution, authentication, permissions, and conversation storage. These components determine whether a coding assistant acts in the intended folder, uses the intended account, and preserves an approved plan. Better reasoning cannot compensate for execution in the wrong directory or a permission rule that fails to apply.
That distinction also provides a useful comparison with OpenAI’s Codex, another coding agent whose surrounding software manages tools and execution controls. This update supplies no comparative performance results. Instead, it gives developers and administrators concrete reasons to reassess how their existing Claude Code workflows behave.
Claude Code v2.1.325 Changes More Than Response Decoding
The security fix leads the announcement, but permission changes give this release broader operational consequences.
Anthropic published the release at 20:40 UTC on November 13, 2026. Its version-specific changes identify a security issue involving zstd, a compression format used to reduce data size. The company says it fixed how compressed responses are decoded.
The announcement does not identify a CVE, a standardized identifier for a publicly disclosed vulnerability. It also does not specify severity, exploitation requirements, or evidence of active exploitation. Those omissions limit what readers can conclude about the response-decoding issue.
Calling the fix evidence of a particular attack would go beyond the published information. So would assuming that every earlier installation faced the same exposure. The documented conclusion is narrower: Anthropic identified a security problem in that decoding path and reports fixing it.
The release also addresses permission enforcement across shell commands, repository settings, background agents, and externally visible artifacts. Artifacts are content objects, such as pages, that an agent can create or modify. Several fixes concern situations where approval or denial rules did not produce their intended behavior.
These entries deserve attention independently of the compression issue. Anthropic reports fixing gaps involving compound Bash commands, which combine multiple shell operations. It also reports a case where an allow rule could execute a different program from the one named.
Other fixes involve relative file rules after a directory change and approvals received after a restart. The common concern is whether an authorization decision remains attached to the action and location it was intended to cover. The changelog documents corrections, without establishing that every affected case shared one underlying cause.
The most consequential default change concerns sandboxed Bash commands. A sandbox is an execution environment that restricts access to resources such as files and network destinations. These commands now follow the normal permission process by default.
Teams seeking the previous automatic behavior must explicitly configure sandbox.autoAllowBashIfSandboxed: true. Where managed settings enable sandboxing, Anthropic says that setting belongs in the highest-ranking managed source. A command remaining inside a sandbox therefore does not necessarily mean it will run without a prompt.
That change separates containment from authorization. Containment limits what a command can reach; authorization determines whether the command should run. Treating those controls separately creates additional decisions for administrators, particularly when workflows depend on unattended execution.
The release remains a Claude Code application update, rather than an announcement of a new foundation model. Its value rests on corrected behavior across existing workflows. The relevant question is whether those workflows now respect their intended boundaries while still completing useful work.
Managed Workflows Face New Approval Decisions
Administrators and automation owners face the clearest immediate pressure because some existing jobs can now pause for approval.
Anthropic changed scheduled tasks and /loop, its recurring task mechanism, to ask before invoking skills that declare tool grants or hooks. A skill packages reusable instructions, while a hook runs configured behavior at a particular event. The documented unattended alternative is an explicit Skill(name) permission allowance.
This change matters when a recurring workflow depends on capabilities bundled inside a skill. A previously quiet job can now reach a legitimate approval boundary. Its pause does not automatically indicate a model failure, a network problem, or a broken scheduler.
Desktop, editor, and software integration hosts also receive a change to persistent approval behavior. Anthropic says a rule alone is offered only when saving it would authorize that command again. The release explicitly warns that a scheduled task relying on an earlier allowance can now wait for input.
Consider an illustrative nightly repository review. The task starts successfully, gathers context, and then invokes a skill containing declared tool access. If that invocation now requires approval, the job can stop progressing despite having launched correctly.
The operational implication is specific: task startup and task completion need separate observation. A successful launch provides little evidence about later permission decisions. Teams evaluating an upgrade should examine where work pauses and which authorized action the pause concerns.
Organization policy messages also become more actionable. When a policy blocks a request, Claude Code adds guidance to use /rewind and change the message, or use /clear. The former revisits conversation history; the latter starts a fresh conversational context.
These instructions describe ways to continue working within the policy. They do not promise that rewording a prohibited request makes it permissible. The improvement gives users a recovery path without presenting the organizational restriction as a transient technical error.
Several other entries correct statements that previously attributed failures to administrators without sufficient evidence. The model selector no longer claims organizational restrictions when Claude Code cannot check them, including when network access is unavailable. Related changes improve messages around fast mode and sign-in conditions.
That distinction affects support work. A verified policy denial calls for a different response than an unavailable policy check. More accurate messages help users direct a problem toward the appropriate owner, although the release provides no measured support-time reduction.
Authentication changes create another reason to review background behavior. Anthropic reports fixing background sessions started through a Claude account that used an API key or token inherited from the launching shell. The release says those credentials could also determine billing.
The announcement separately describes refusing more cases where sessions moved to another sign-in in the background. These changes concern account identity as well as command execution. For organizations, a completed task still needs to belong to the intended account and administrative context.
These entries do not establish a billing incident at any particular customer. They identify corrected product behavior. Teams whose services launch background sessions from credential-rich shells have a concrete scenario to examine, rather than a reason to assume every job was affected.
Permission Boundaries Reach Beyond the Main Project
The update treats working directories, repository metadata, and agent configuration as boundaries that need explicit protection.
Anthropic’s sandbox configuration distinguishes execution restrictions from permission decisions. That distinction helps explain why this release includes both tighter protections and additional approval requirements. A sandbox can remain enabled while the rules for running commands inside it change.
The release extends protected sandbox paths into added directories. Sandboxed commands now refuse writes to .claude settings, skills, agents, commands, and hooks there, matching protection in the launch folder. Those files can influence future agent behavior, making their treatment consequential.
This creates a practical conflict during ordinary source-control operations. A branch switch or stash operation can include configuration files alongside application code. Blocking changes to protected configuration can therefore interrupt a command whose broader purpose appears routine.
Anthropic adds guidance for Claude when sandboxed git checkout, git stash, or an autostash pull cannot update protected files. Autostash temporarily sets aside local changes during a pull. The added note addresses the agent’s understanding of the failure, without removing the protection.
The distinction matters during recovery. Repeating the same blocked operation does not resolve a protected-path conflict. An accurate explanation gives the assistant and user a better basis for deciding what authorized action should follow.
Git compatibility improves separately. A worktree is an additional working checkout connected to a repository. Anthropic reports fixing /security-review and inline Git commands in skills for worktrees of bare repositories.
A bare repository stores Git data without the usual checked-out project files. The same fix covers --separate-git-dir checkouts and Android repo checkouts, which organize repository data differently. These are compatibility corrections for documented layouts, rather than new source-control features.
The release also closes cases where the Bash sandbox did not fully protect worktree settings, hooks, or shared Git storage. Shared metadata makes isolation more complicated than protecting the visible project directory alone. A command’s effects can extend into repository structures used by another checkout.
Another change prevents silent movement after a working directory disappears. Previously, shell commands could run in the launch directory instead. Anthropic says the command now fails with a message and can be issued again.
For an agent interpreting relative paths, that behavior is significant. The same command text can target different files depending on its location. Failing clearly gives the user a chance to restore the intended location before work continues.
OpenAI describes comparable infrastructure concerns in its discussion of the Codex agent harness. That surrounding software manages conversation state, tools, and configured execution controls. The architectural comparison explains why these concerns extend beyond any particular model’s coding ability.
It does not establish that Claude Code and Codex enforce identical boundaries. Nor does this release support a ranking of their security. The useful comparison concerns evaluation: both reasoning quality and the machinery translating decisions into actions deserve scrutiny.
For a team assessing coding agents, a generated patch is only part of the result. Its location, permission history, account identity, and preservation of protected settings also matter. Claude Code v2.1.325 makes those operational details unusually visible through its documented changes.
Session and Subagent Fixes Address Different Reliability Gaps
Preserving work and applying the intended model configuration are separate requirements, and this release addresses both.
Anthropic reports fixing recent conversation turns that sometimes disappeared when users resumed sessions after an unclean exit. An unclean exit means the application stopped without completing its normal shutdown. The announcement does not quantify affected sessions or claim universal recovery of previously missing content.
The distinction between preventing loss and recovering old loss matters. A repair to future session handling does not establish that earlier missing turns can be reconstructed. Teams should avoid reading a stronger recovery guarantee into the entry.
Approved plans receive a related persistence correction. If the configured plansDirectory cannot accept a plan edit, Claude Code now saves it in the default plans folder. A warning identifies the destination and explains the fallback.
That behavior makes a failed save location visible while preserving the approved text elsewhere. It also creates a distinction between the requested storage location and the actual one. Workflows that collect plan files need to account for the warning and destination.
The update also fixes sessions on many third-party gateways and Foundry that repeatedly stopped with context-length errors instead of compacting. Compaction condenses conversation history to fit the model’s available context. A gateway is an intermediary service through which model requests pass.
The reported correction concerns continuing a session when its accumulated content reaches a limit. It does not establish that every condensed conversation preserves every instruction or detail. Successful compaction and faithful retention remain different questions.
Subagents, specialized assistants that perform delegated work, receive several configuration fixes. Anthropic’s subagent model controls explain that these assistants can operate with their own model choices. This makes correct configuration important when the parent session uses different settings.
The release fixes subagents with their own models being moved to the planning model under the opusplan or haiku settings. It also fixes file-read and tool-result sizing that used the main session’s model instead of the subagent’s model.
Those tool results can arrive through MCP, the Model Context Protocol used to connect agents with external tools and services. Different model limits make incorrect sizing relevant to what a delegated assistant can receive. The release does not quantify the previous effect on task quality.
Anthropic also improves handling when a subagent reaches its output token limit, the maximum response length measured in model text units. The assistant now receives a request to provide the complete answer more briefly. Previously, it could return only the ending.
These changes address delegation mechanics rather than introducing a new agent capability. A workflow still needs to establish that delegated findings are complete and useful. The changelog offers no benchmark showing a particular improvement in accuracy, speed, or resource consumption.
Smaller interface changes address the same broader need for understandable state. After Claude reads or edits a Jupyter notebook, an interactive document containing executable cells, integrated terminals can show a cell-rerun tip. Anthropic says this appears in VS Code and similar terminals.
That reminder distinguishes editing notebook code from executing it. An altered cell does not itself demonstrate that displayed results reflect the new code. The tip supports that distinction without claiming to validate notebook outputs.
Together, these fixes describe several ways a session can appear successful while leaving an incomplete record or mismatched configuration. Their significance lies in the specific failure modes. Treating them as proof of general reliability would exceed the evidence.
What Remains Unclear, and What to Watch Next
The release documents substantial corrections, but it does not measure how reliably those corrections hold across customer environments.
The response-decoding disclosure remains limited. Anthropic identifies the affected operation and says it fixed a security issue. Without additional technical detail, readers cannot determine the attack prerequisites or assign a defensible severity from the release alone.
The wider permission entries also require careful interpretation. Their number does not establish an exploitation rate or a comparative security score. Several concern different surfaces, including local commands, cloud sessions, artifact publication, and worktree metadata.
The announcement provides no independent validation of these fixes. It also supplies no customer incident count or measured reduction in failures. The strongest supported reading is that Anthropic shipped documented corrections and changed several defaults with practical consequences.
The first signal to watch is further security disclosure. A technical advisory describing affected versions, prerequisites, or mitigations would clarify the response-decoding issue. It would strengthen the basis for prioritizing exposure, while potentially narrowing or expanding the initially understood scope.
Further release notes about the same permission paths would also matter. A follow-up correction would show that a documented boundary needed additional work. Conversely, an absence of follow-up entries would remain insufficient to prove that no problems persisted.
The second signal is actual permission behavior in representative managed workflows. The relevant evidence is whether expected jobs finish, pause at intended approval points, and refuse prohibited actions. Merely checking that the application starts would leave the central operational changes untested.
For example, an unattended skill invocation should follow the organization’s explicit permission configuration. A sandboxed command should not gain access to protected settings through an added directory. Successful behavior in those cases would support the release’s practical value for that environment.
Unexpected approval prompts need interpretation rather than automatic dismissal. Some are the intended result of the new default. Others can reveal a mismatch between a team’s saved permissions and its current execution setup.
The third signal is session continuity under the configurations named in the release. Interrupted sessions, configured plan folders, third-party gateways, and separately configured subagents provide concrete areas to examine. Evidence from an unrelated default setup would not settle behavior in those cases.
For session recovery, the useful question is whether the latest intended conversation state returns after interruption. For plans, it is whether the approved content survives and its actual storage location remains clear. For subagents, it is whether the configured model and usable results match expectations.
Teams can preserve those observations alongside their technical documentation. Keeping failure examples, configuration decisions, and verified outcomes in a searchable knowledge base helps future reviewers distinguish established behavior from assumptions. That record complements the coding agent’s own conversation history.
Anthropic’s release gives developers a concrete reason to review the controls surrounding their AI-assisted work. Start with a workflow that matters, then examine its permission decisions, execution location, account identity, and saved results. Compare those observations with the specific corrections documented for Claude Code v2.1.325. Record where stronger approvals intentionally interrupt automation, and distinguish those pauses from unresolved failures. Preserve the outcome outside a single conversation so another developer can understand what was verified. For teams choosing when to adopt an update, that evidence supports a clearer decision than either a successful launch or a long changelog. Which boundary in your workflow still depends on an assumption you have never checked?



