On October 21, a poisoned litellm==1.107.0 sat on PyPI for under two hours1 and was downloaded roughly 49,000 times, by Wiz’s count.2 Its payload did not prompt-inject anything. It ran at interpreter start and read SSH keys, shell history, cloud and container credentials, “the environment variables of every readable process,” and the credential files and configs of Claude, Codex, Gemini, Cursor, Windsurf, and Copilot, among others.3 And the way the attackers got into the release pipeline, according to LiteLLM’s completed root-cause analysis, was an organization owner’s GitHub CLI OAuth token sitting in an internal agent platform’s configuration. The platform’s API returned it to anyone holding the master key, which had leaked in a public commit in May and was never revoked.4
That is one incident with two credential failures and zero model failures. Nobody talked a model out of its instructions. A secret was sitting where code could read it, twice: once as a file on disk, once as a field in an agent’s configuration.
My argument is that most of the effort going into prompt-injection defense is aimed at the wrong layer. “Can the model be tricked” is an open research problem with no clean answer. “Can the process the agent runs code in read the secret” is a yes-or-no question you can answer by inspection. The pattern that makes the answer no already exists. It is shipping in brokers, in gateways, and in at least one agent runtime. What it is not, anywhere I looked, is the default.
Two reachability failures and no model failure
Start with the stealer. According to the researcher who decoded it in LiteLLM’s disclosure thread, each 1.107.0 wheel carried a cook.pth, which pip dropped into the site-packages root so the stealer ran on every start of that interpreter, and a sitecustomize.py holding a dormant loader for remote code.5 The decoded module list is in the thread: ssh_keys, shell_hist, cloud_creds, container_creds, cicd_secrets, env_vars, and one named ai_devtools.6 On a LiteLLM proxy with DATABASE_URL set, it also dumped the two database tables holding every upstream provider key the proxy stored.7
Concede the obvious. This is ordinary malware. A .pth infostealer would have worked in 2019, and nothing about reading ~/.aws is specific to agents. What is new is the target list. A stealer with a module named ai_devtools is a stealer whose authors have noticed where the keys moved. They moved into the config files and session logs of coding agents. GitGuardian’s September analysis of that endpoint problem reported 24,008 unique secrets in public MCP configuration files, 2,117 of them still valid, and a 3.2% secret-leak rate in public commits assisted by Claude Code against a 1.5% baseline.8 GitGuardian sorts the leaks three ways: stored in agent or MCP config, copied out of a vault into a local file, or recorded in prompts, command output, and session history.9 All three are the same failure. The secret was somewhere the agent’s process could read it.
The second failure matters more, and it is the one that moved while I was writing. LiteLLM’s first incident post, on October 21, said the attacker had obtained GitHub CLI OAuth tokens from the logs of an internal LiteLLM deployment that was exposed to the public internet,10 and gave a one-sentence account of how they got into those logs:
After a thorough investigation, we have observed specific situations in which Claude Code will embed a user’s GitHub CLI OAuth token into a response body, and in situations where response bodies are logged, these tokens may appear in LiteLLM logs.11
As of October 28 I can find no statement from Anthropic about that sentence either way. LiteLLM has moved past it. Its completed attack-chain analysis, published October 23, calls itself “the canonical source of truth,” says some beliefs in the earlier updates “were later proven to be inaccurate or incomplete,” and does not mention Claude Code. It says that the “first post suspected the tokens were taken from the logs of an internal LiteLLM deployment” and that the request logs show the first token “came from a different service.”12
That service was the LiteLLM Agent Platform, an internal tool for self-hosted agents that open pull requests. On May 27 a configuration file holding its master key was committed to a public repository, force-pushed over, and never revoked. On October 10 the attacker used that key to call the platform’s API, and a request for the list of managed agents returned every agent’s configuration, including one owner’s GitHub CLI OAuth token, which “was not associated with a service account and was not scoped down.”13 With the owner’s token the attacker pushed throwaway branches whose workflows posted two more maintainers’ tokens, stored as CI secrets, to an attacker-controlled endpoint. On October 21 the attacker used those three tokens to override branch protection, point the release workflow at a lookalike Berri-AI/litellm, approve the production gate, and delete the run records.14
Notice what left the story and what did not. The agent that supposedly repeated a token is gone. What remains is an agent platform that kept a human owner’s unscoped token in an agent’s configuration, where any caller with the master key could read it, and a CI system that kept two more in secrets any pushed workflow could read. GitGuardian’s first pathway, a secret stored in agent config, is the one that opened the pipeline. The model did nothing. The secret was within reach.
The guidance is looking one layer up
Simon Willison’s lethal trifecta has been the clearest framing of this since June 2025: an agent is dangerous when it has access to private data, is exposed to untrusted content, and can communicate externally. The only safe move, in that framing, is to avoid the combination.15 Almost all of the published effort I can find goes at the untrusted-content leg: classifiers, guardrails, model training. I went looking for current guidance that treats the private-data leg as its own problem, at the layer where both LiteLLM failures happened, and mostly did not find it.
The OWASP Top 10 for Agentic Applications, published December 2025, files credential misuse under ASI03, “Identity and Privilege Abuse,” and counts “cached credentials” and “API keys, OAuth tokens, delegated user sessions” as part of what an agent’s identity means.16 Its mitigations are short-lived, narrowly scoped tokens, per-session sandboxes with separated memory, per-action authorization, and human approval for escalation.17 All of that is sensible, in my view. None of the nine items asks whether the agent’s process can read the token at all. Shorten it, scope it, wrap a sandbox around the session: the token is still inside the sandbox. OWASP’s own Q3 2026 exploit roundup, published October 8, puts the ASI03 tag on seven of its entries and includes “Keep platform tokens outside containers” among its per-incident fixes.18 That is the closest the project gets to the layer I mean.
CoSAI’s zero-trust paper, dated September 4, gets closer in architecture. It draws an “AI Threat Boundary” around the model, says everything inside it “is non-deterministic and must be treated as untrusted,” and puts credential handling in a deterministic orchestration layer outside that boundary, where static secrets “are pulled just-in-time from the Vault.”19 I think that is the right diagram. The problem is that in the harnesses people run, the orchestration layer and the agent’s shell share a process, an environment, and a filesystem. The diagram assumes a separation the runtimes do not have.
Unit 42 demonstrated that gap in September against AWS’s AgentCore. A hidden HTML comment in a support ticket got the agent to fetch and run a script, and the shell tool ran as root inside the harness. The script read /proc/1/mem and pulled a 1,034-byte identity JWT for the mcp-service account out of the harness’s heap. That token worked from an outside machine with no AWS credentials at all.20 AWS closed the report as informative under its shared-responsibility model and pointed at allowedTools scoping and egress filtering as customer-side controls.21 I think that answer is defensible. Egress filtering and tool scoping do close that path. The question is why the configuration where the secret is unreachable is the one the customer has to go and build.
The Australian Signals Directorate’s harness guidance, published September 11, comes closest to naming the layer. ASD says “no fully reliable technical mitigation currently exists” for prompt injection, so mitigations “must instead be applied in the harness, by controlling what an agent can access and what actions it is permitted to perform.”22 That is, I think, the right sentence. What follows it is least privilege, human approval, and logging. It stops one step short of the specific question: can the code the agent runs read the credential?
The pattern exists. The runtimes made it opt-in.
Secretless brokering is not new, and I am not going to pretend it is. HashiCorp’s Vault Proxy “acts as a proxy for Vault’s API, optionally using (or forcing the use of) the auto-auth token,” so the application behind it never handles the token itself.23 Cloudflare’s AI Gateway keeps provider keys in its Secrets Store and “substitutes the stored provider key only when the provider authorization header is absent,” so a client can call a model API without carrying the key.24 A SANS Institute author filed an IETF draft in March, Credential Broker for Agents, whose first model is a proxy gateway and whose premise is that “agents never hold real long-lived credentials.”25 Gravitee published the same pattern for MCP in July.26 So did agentgateway’s Christian Posta, in one line: “the agent makes its call without a credential, and something else on the path attaches the real one before the request reaches its destination.”27
So the pattern is standard in infrastructure. What interested me is that at least one agent runtime has built it against the actual blast radius, the sandbox the agent runs code in. Omnigent’s “Secretless Credential Proxy” design, marked implemented, landed on June 16 and was extended on July 25.28 Its statement of the problem is blunt: putting the real token in the sandbox “defeats much of the point of the sandbox,” because any code the agent runs can read it out of the environment and send it somewhere.29 The default is swap-on-access. The tool sends its request with no Authorization header, and the mandatory egress proxy in the unsandboxed parent attaches the credential on the way out, so that “nothing credential-shaped enters the sandbox.”30
The exception is the instructive part. Some clients refuse to send a request without a local credential. The gh client stops before it touches the network. For those, the parent mints a single-use placeholder prefixed oa_cred_, bound to one host, and swaps it at the proxy; a placeholder presented to any other host gets HTTP 403.31 Read that against LiteLLM. A placeholder sitting in an agent platform’s configuration is worth nothing to the attacker who pulls it out through the API.
Coverage is narrow, and the design says so. It requires egress rules and a hard-isolating backend, linux_bwrap or darwin_seatbelt, and it covers HTTP bearer and basic auth only, with SSH remotes out of scope.32 Windows gets a Job Object that contains the process tree but “does not isolate the filesystem or network,” and the egress proxy is not available there.33 One project, two platforms, one protocol.
Here is the part that changed my read while I was writing. Claude Code itself ships nearly the same mechanism. Its sandbox documentation describes masking a credential so that sandboxed commands see “a per-session placeholder called the sentinel” while the sandbox proxy “substitutes the real value on outbound requests to hosts you allow,” and its worked example is GH_TOKEN substituted only on requests to api.github.com.34 That is the Omnigent design, down to the placeholder. Then come the defaults. The sandbox “is off by default.”35 With it on, sandboxed commands can still read “most of the machine, including credential files such as ~/.ssh and ~/.aws/credentials,” and they inherit Claude Code’s environment, “including any secrets in its environment.”36 “There is no built-in credential deny list, so only the files and variables you list are restricted.”37 Mods, the new in-process extension mechanism, “aren’t sandboxed” and can “read your secrets: environment variables and settings files, including an API key you keep in either.”38 File tools, MCP servers, and hooks run outside the sandbox.39
I want to be fair to Anthropic here. I think that documentation is unusually candid, and the masking feature is real engineering. But the shape is the same as AWS’s. The secure configuration exists, and it is the one you have to go and turn on.
What I think survives
This is opinion, stated plainly.
The LiteLLM details changed once while I was writing. Between the first post and the canonical one, the token that opened the pipeline moved from a gateway’s logs to an agent platform’s configuration. The argument did not move. Both failures were a secret sitting where code could read it, and a better injection classifier would not have prevented either.
So the question I think a team should put to any agent runtime is not “how well does it resist injection.” It is two questions. Can the process the agent runs code in read the credential? Does anything the agent can see end up in a log? Both are checkable in an afternoon. Open a shell inside the sandbox and print the environment. Read the gateway’s logging config and look for response bodies. If the answer to the first question is yes, the model’s resistance to being talked into things is the only control between the attacker and the secret. That is the one control nobody has made reliable.
My read is that the secretless pattern wins, because it is binary, inspectable, and already built by the brokers and gateways. What I am less sure of is how long the defaults take to follow. Every runtime I looked at ships the credential in reach and the broker as an option. Until that flips, the quickstart is the vulnerability.
Notes
Footnotes
-
BerriAI, “Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised,” docs.litellm.ai, October 21, 2026 (last updated 20:30 UTC that day). States the package was available on PyPI from 11:39:51 to 13:23 UTC. Opened October 23, 2026; re-opened later that day, when the page carried an added note dated October 23, 21:50 UTC pointing to the completed attack-chain analysis; re-opened again October 28, 2026, unchanged. https://docs.litellm.ai/blog/security-incident-oct-21-2026 ↩
-
Benjamin Read and Rami McCarthy, “LiteLLM Compromised, Steals Credentials,” Wiz, October 21, 2026. Vendor research blog; source for the “approximately 49,000 downloads” figure. Opened October 23, 2026; re-opened October 28, 2026. https://www.wiz.io/blog/litellm-compromised-steals-credentials ↩
-
LidorMachluf, “[SECURITY] litellm 1.107.0 compromised,” BerriAI/litellm issue #47826, opened October 21, 2026, 12:06 UTC. The issue body lists what the stealer collects, including
~/.ssh/*, shell histories, cloud and container credentials, “the environment variables of every readable process,” and “AI tool credentials and configs: Claude, Codex, Gemini, Qwen, Cursor, Windsurf, Kiro, VS Code MCP and Copilot.” Opened October 23, 2026 via the GitHub API; re-opened October 28, 2026. https://github.com/BerriAI/litellm/issues/47826 ↩ -
Oliver Jensen and Mateo Wang, “The attack chain behind the v1.107.0 LiteLLM release,” docs.litellm.ai, October 23, 2026. The master key of the LiteLLM Agent Platform was committed to the public
LiteLLM-Labs/lite-harnessrepository on May 27, 2026 and not revoked; on October 10 a call to the platform’s API with that key returned the GitHub CLI OAuth token of one of the organization’s two GitHub owners. Opened October 23, 2026; re-opened October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain ↩ -
Issue #47826 body, October 21, 2026.
cook.pth,cook.py, andsitecustomize.pyare installed to the site-packages root; the.pthimports the stealer on every interpreter start;sitecustomize.pyforks a loader that polls for remote code. Opened October 23, 2026; re-opened October 28, 2026. https://github.com/BerriAI/litellm/issues/47826 ↩ -
LidorMachluf, “Decoded payload,” comment on issue #47826, October 21, 2026, 12:57 UTC. The decoded source sets
PC_MODULESto a list that includesssh_keys,shell_hist,cloud_creds,container_creds,cicd_secrets,ai_devtools, andenv_vars. Opened October 23, 2026; re-opened October 28, 2026. https://github.com/BerriAI/litellm/issues/47826 ↩ -
Issue #47826 body, October 21, 2026: “if
DATABASE_URLis set, it spawns prisma and dumpsLiteLLM_CredentialsTableandLiteLLM_ProxyModelTable, i.e. every upstream provider key the proxy stores.” Opened October 23, 2026; re-opened October 28, 2026. https://github.com/BerriAI/litellm/issues/47826 ↩ -
GitGuardian, “AI Coding Agents Are Leaking Credentials: Cursor, Claude Code, Copilot, and MCP,” GitGuardian Blog, September 25, 2026. Vendor research blog. The 24,008 and 2,117 figures and the 3.2% versus 1.5% leak rates are attributed there to GitGuardian’s State of Secrets Sprawl 2026 report. Opened October 23, 2026; re-opened October 28, 2026. https://blog.gitguardian.com/ai-coding-agents-credential-security/ ↩
-
GitGuardian, same post, September 25, 2026. The three pathways are labelled stored, copied, and recorded. Opened October 23, 2026; re-opened October 28, 2026. https://blog.gitguardian.com/ai-coding-agents-credential-security/ ↩
-
BerriAI, “Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised,” October 21, 2026 (last updated 20:30 UTC that day). Describes “an unexpected, innocuously named user account in our internal LiteLLM deployment, which was exposed to the public internet,” and states the belief that the attacker “obtained the GitHub CLI OAuth tokens from its logs.” Opened October 23, 2026; the sentence was still on the page when re-opened later that day and on October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026 ↩
-
BerriAI, same post, October 21, 2026. Quoted verbatim. Opened October 23, 2026, and still present verbatim when re-opened later that day and on October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026 ↩
-
Oliver Jensen and Mateo Wang, “The attack chain behind the v1.107.0 LiteLLM release,” docs.litellm.ai, October 23, 2026. Quoted phrases verbatim: “This post is the canonical source of truth”; “some assumptions and beliefs in those updates were later proven to be inaccurate or incomplete”; and, under the compromised-credentials table, “Our first post suspected the tokens were taken from the logs of an internal LiteLLM deployment. The request logs show the first token came from a different service, described in step 1.” The post does not contain the words “Claude Code.” Opened October 23, 2026; re-opened October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain ↩
-
Same post, “Narrative,” “GitHub Owner Access,” and “Step 1” sections, October 23, 2026. The Agent Platform is described as “an experimental project for self-hosted cloud agents to file PRs to GitHub repositories”; the May 27 commit, the failed revocation, and the October 10 request
GET /api/v1/managed_agents/agents?limit=100that “returned the configuration of every managed agent” are described there. Quoted phrase verbatim. Opened October 23, 2026; re-opened October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain ↩ -
Same post, “Step 2” through “Step 7,” October 23, 2026. Workflow runs on throwaway branches posted
GH_TOKENandRELEASE_PLEASE_TOKEN, each a maintainer’s GitHub CLI OAuth token, to a webhook.site endpoint on October 10; on October 21 the owner’s token overrode branch protection and dispatched the release with the source set to “the lookalike Berri-AI/litellm,” the second maintainer’s token approved the production gate, and the third’s deleted the run records. Opened October 23, 2026; re-opened October 28, 2026. https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain ↩ -
Simon Willison, “The lethal trifecta for AI agents: private data, untrusted content, and external communication,” simonwillison.net, June 16, 2025. Opened October 23, 2026; re-opened October 28, 2026. https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ ↩
-
OWASP Gen AI Security Project, “OWASP Top 10 for Agentic Applications 2026,” Version 2026, December 2025, ASI03 “Identity and Privilege Abuse,” pp. 15–16. Identity “includes both the agent’s assigned persona and any authentication material (API keys, OAuth tokens, delegated user sessions) that represent it.” PDF opened October 23, 2026 from the resource page; re-opened October 28, 2026. https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ ↩
-
Same document, ASI03 “Prevention and Mitigation Guidelines,” nine items, p. 17. Opened October 23, 2026; re-opened October 28, 2026. https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ ↩
-
OWASP Gen AI Security Project, “GenAI and Agentic AI Exploit Roundup Q3 2026,” October 8, 2026. The label “ASI03 Identity and Privilege Abuse” appears on seven entries; the mitigation list for the Modal CyberGym entry includes “keep platform tokens outside containers.” Opened October 23, 2026; re-opened October 28, 2026. https://genai.owasp.org/2026/10/08/genai-and-agentic-ai-exploit-roundup-q3-2026/ ↩
-
Coalition for Secure AI, Workstream 2, “Zero Trust for AI Systems: Why Authorization Can’t Live Inside the Model,” Version 1.0, dated September 4, 2026 and posted September 16, 2026. Section 3.2.2 describes the AI Threat Boundary and the orchestration layer; static secrets “are pulled just-in-time from the Vault.” PDF opened October 23, 2026; re-opened October 28, 2026. https://www.coalitionforsecureai.org/wp-content/uploads/2026/09/Zero-Trust-for-AI-Systems.pdf ↩
-
Niv Rabin, “A Vault with a Heap-View: The Uncomfortable Space Between AgentCore Harness and Identity,” Palo Alto Networks Unit 42, September 18, 2026. Vendor research blog. Opened October 23, 2026; re-opened October 28, 2026. https://unit42.paloaltonetworks.com/securing-aws-agentcore-harness-credentials/ ↩
-
Same Unit 42 post, September 18, 2026: AWS “closed the report as informative under the AgentCore shared responsibility model, citing allowedTools scoping and egress filtering as customer-side controls.” Opened October 23, 2026; re-opened October 28, 2026. https://unit42.paloaltonetworks.com/securing-aws-agentcore-harness-credentials/ ↩
-
Australian Signals Directorate, “Agentic AI harnesses: the layer above the model,” cyber.gov.au, first published September 11, 2026. Quoted verbatim. The live page timed out on October 23 and again on October 28, 2026, so the wording was confirmed against the Internet Archive’s September 23, 2026 snapshot (https://web.archive.org/web/20260923024056/https://www.cyber.gov.au/business-government/secure-design/artificial-intelligence/agentic-ai-harnesses). Reported by Juha Saarinen, “ASD says prompt injection in AI cannot be fixed,” iTnews, September 21, 2026. https://www.cyber.gov.au/business-government/secure-design/artificial-intelligence/agentic-ai-harnesses ↩
-
HashiCorp, “What is Vault Proxy?,” Vault documentation, undated (accessed October 23 and October 28, 2026). Quoted verbatim. https://developer.hashicorp.com/vault/docs/agent-and-proxy/proxy ↩
-
Cloudflare, “Bring your own keys (BYOK),” AI Gateway documentation, undated (accessed October 23 and October 28, 2026). Quoted verbatim. https://developers.cloudflare.com/ai-gateway/configuration/bring-your-own-keys/ ↩
-
Kenneth G. Hartman (SANS Institute), “Credential Broker for Agents (CB4A),” Internet-Draft draft-hartman-credential-broker-4-agents-00, March 29, 2026. The draft expired September 30, 2026 and is cited as a design statement, not a standard. Opened October 23, 2026; re-opened October 28, 2026. https://datatracker.ietf.org/doc/html/draft-hartman-credential-broker-4-agents-00 ↩
-
Prachi Jamdade, “MCP Credential Brokering: Keep API Keys From AI Agents,” Gravitee, July 31, 2026. Vendor blog. Opened October 23, 2026; re-opened October 28, 2026. https://www.gravitee.io/blog/mcp-credential-brokering-keep-api-keys-from-ai-agents ↩
-
Christian Posta, “Credential Injection Patterns for AI Agents,” agentgateway blog, July 27, 2026 (the date shown on the page; a follow-up part is dated August 3). Vendor blog. Quoted verbatim. Opened October 23, 2026; re-opened October 28, 2026. https://agentgateway.dev/blog/2026-07-27-credential-injection-ai-agent-egress-cb4a/ ↩
-
omnigent-ai/omnigent,
designs/SANDBOX_CREDENTIAL_PROXY.md, “Secretless Credential Proxy,” status IMPLEMENTED. Commit history for the file via the GitHub API: “feat(sandbox): secretless credential_proxy for egress (bearer + basic)” (#236), June 16, 2026, and “Add databricks_cli secretless credential proxy type” (#3080), July 25, 2026. Opened October 23, 2026; re-opened October 28, 2026. https://github.com/omnigent-ai/omnigent/blob/main/designs/SANDBOX_CREDENTIAL_PROXY.md ↩ -
Same design document. Quoted phrase verbatim; the document’s full sentence names
os.environand~/.gitconfigas where the token would be read from. Opened October 23, 2026; re-opened October 28, 2026. https://github.com/omnigent-ai/omnigent/blob/main/designs/SANDBOX_CREDENTIAL_PROXY.md ↩ -
Same design document, swap-on-access section. Quoted phrase verbatim. Opened October 23, 2026; re-opened October 28, 2026. https://github.com/omnigent-ai/omnigent/blob/main/designs/SANDBOX_CREDENTIAL_PROXY.md ↩
-
Same design document: the parent “mints a random, single-use placeholder prefixed
oa_cred_”; “A placeholder presented for a host it is not bound to (an exfiltration attempt) is rejected with HTTP 403.” Opened October 23, 2026; re-opened October 28, 2026. https://github.com/omnigent-ai/omnigent/blob/main/designs/SANDBOX_CREDENTIAL_PROXY.md ↩ -
Same design document: requires
egress_rulesand a hard-isolating backend (linux_bwrapordarwin_seatbelt); scope is “HTTP(S) Bearer/Basic only. SSH-based git remotes are out of scope.” Opened October 23, 2026; re-opened October 28, 2026. https://github.com/omnigent-ai/omnigent/blob/main/designs/SANDBOX_CREDENTIAL_PROXY.md ↩ -
omnigent-ai/omnigent,
README.md, undated (accessed October 23 and October 28, 2026). The Windows Job Object backend “does not isolate the filesystem or network”; the L7 egress proxy is “not available on Windows.” https://github.com/omnigent-ai/omnigent/blob/main/README.md ↩ -
Anthropic, “Configure the sandboxed Bash tool,” Claude Code documentation, section “Mask credentials,” undated (accessed October 23 and October 28, 2026). Quoted verbatim; the
GH_TOKENexample setsinjectHoststoapi.github.com. https://code.claude.com/docs/en/sandboxing ↩ -
Same page, section “What the sandbox restricts”: “The sandbox is off by default.” Accessed October 23 and October 28, 2026. https://code.claude.com/docs/en/sandboxing ↩
-
Same page, default-access table: reads cover “Most of the machine, including credential files such as
~/.sshand~/.aws/credentials”; environment variables are “Inherited from Claude Code, including any secrets in its environment.” Accessed October 23 and October 28, 2026. https://code.claude.com/docs/en/sandboxing ↩ -
Same page, section “Protect credentials.” Quoted verbatim. Accessed October 23 and October 28, 2026. https://code.claude.com/docs/en/sandboxing ↩
-
Anthropic, “Mods overview,” Claude Code documentation, section “What a mod can reach,” undated (accessed October 23 and October 28, 2026). Quoted verbatim. The page says mods are on by default from Claude Code v2.1.287. https://code.claude.com/docs/en/plugins/mods/overview ↩
-
Anthropic, “Configure the sandboxed Bash tool,” section “What runs outside the sandbox”: built-in file and web tools, command hooks, local MCP servers, and other helper processes “run with your full access.” Accessed October 23 and October 28, 2026. https://code.claude.com/docs/en/sandboxing ↩


