Executive reading · ~60 seconds

The AG2 advisory demonstrates that a command list and a read-only mode do not form a sandbox when the filter and shell assign different semantics to the same input.

A team may restrict a tool to commands considered safe and still pass the shell input that it interprets differently. The AG2 advisory shows a concrete flaw at this boundary: policy validated one representation of the input, while execution obeyed the interpreter's full grammar.

Status: Evidence Brief approved for publication. Editorial series: Papers we read for you. The text omits payloads, targets, and operational exploitation instructions.

The verifiable fact

O GHSA-2h4w-29cv-2cmv rates a flaw in the following component as High: SandboxShellTool. According to the record, the allowlist and read-only modes could accept input containing a separator not covered by the filter and pass it to sh -c, which then interpreted more than one instruction.

The package ag2 is affected from 0.13.2 up to, but not including, 1.1.1; version 1.1.1 is identified as fixed. The record also lists autogen from 0.13.2 to 0.14.1 and, at the editorial cutoff, declares no fixed version for that package. This is version information from the advisory, not a claim that every installation is exposed or has been exploited.

Population, method, and control

The observed population is narrow: two restricted modes of the same tool and the version ranges stated in the record. The researcher analyzed the real filter, repeated its allowlist branch, and compared inputs blocked by the known operator set with the unrecognized variation. The researcher then confirmed that the same interpreter used by the product assigns executable semantics to the accepted variation.

The control matters: operators explicitly covered by the filter were rejected. The evidence is therefore not merely “the shell executes commands”; it demonstrates the divergence between what policy considered a single instruction and what the executor considered a sequence. The record also specifies that the default local backend ran subprocesses on the host, without real container isolation.

There is no denominator of installations, customers, environments, or incidents. At the editorial cutoff, we found no public confirmation of production exploitation or additional independent reproduction. Publication by the repository and the fixed version constitute the maintainer's response, not independent corroboration.

The architecture lesson

A prefix allowlist is an intent control, not an execution boundary. If validation operates on text and execution delegates that text to another interpreter, policy must use the same grammar — or eliminate the intermediate interpretation. In the Forge canon, explicit capability, isolation, negative testing, and execution receipts are complementary controls. [R2-C3]

DecisionMinimal evidenceFailure the test must capture
allowed commandstructured arguments, without ambiguous text reconstructionthe input must not gain additional semantics in the executor
read-only modeallowed operations and resources defined by capabilityan operation outside the capability must fail before process execution
local backendidentity, directory, network, and filesystem actually isolatedthe subprocess must not inherit implicit host access
updatepackage and version actually loaded at runtimeupdating the lockfile without replacing the process does not close the risk
agent useobservable instruction origin, approval, and resultuntrusted content cannot expand capability through inducement

What to do without turning the advisory into an offensive playbook

  1. Inventory ag2 and autogen, including the version actually imported and the backend in use.
  2. Update ag2 to 1.1.1 or later; for autogen, follow the official guidance because the advisory did not declare a fixed version at the cutoff.
  3. Treat readonly=True and command lists as policy, not as host isolation.
  4. Prefer execution with structured arguments and real process, filesystem, identity, and network boundaries.
  5. Add negative tests for control characters, normalization, encodings, and differences among parser, validator, and executor, without storing dangerous payloads in public documentation.
  6. Record the rollout receipt: version, image or environment, configuration, tests, and replaced process.

Limitations and counterpoints

The existence of the flaw does not prove that any application built with AG2 is compromised. Exploitation depends on the vulnerable tool being enabled, content or an actor being able to influence its input, and the backend granting relevant access. Restricted container deployments may reduce impact, although they do not correct the filter's wrong decision.

The advisory includes enough technical detail for maintainer validation, but it is a single record and does not measure prevalence. It also does not support concluding that other shell tools have the same implementation. The defensible conclusion is narrower: in this component and these versions, the contract declared by the restricted modes did not match executor behavior.

Direct sources

  1. GHSA-2h4w-29cv-2cmv, advisory and official project response.
  2. CVE-2026-113663, CVE record.
  3. AG2 1.1.1 on PyPI, version identified as fixed.

Editorial and responsibility note

This text combines facts attributed to sources with the author's analysis and technical proposal. Personal and professional views are not proven facts; versions, conditions, and limits are stated when available. The content is informational and does not replace technical, legal, regulatory, or security assessment. AG2, AutoGen, GitHub, PyPI, and other marks belong to their respective owners; there is no sponsorship or endorsement. No offensive procedure is reproduced. Tech Human and Trustyu operate commercially in digital-product development and governance. Research, structure, and writing received AI assistance; Fernando Parreiras authorized the topic, factual self-review, and publication on November 11, 2026, without additional independent human review.

Research cutoff: November 11, 2026. Editorial status: special Evidence Brief authorized for November 21, 2026 at 09:00 BRT.

Editorial and responsibility note

Research cutoff
Last review
Recorded corrections
No corrections recorded.

The cutoff above applies to the canonical claims. Additional sources and their access dates are identified in the article body.

This article combines cited sources, analysis, and the author's professional experience. Verifiable data and factual statements are linked to their respective sources. Interpretations, hypotheses, projections, recommendations, and opinions represent the author's professional point of view at the time of publication; they do not constitute proven facts, a promise of results, or legal, financial, or technical advice applicable to a specific case. Consult the original sources and qualified professionals before making decisions.

Claims and sources

R2-C3

Least privilege, isolated execution and supply-chain provenance are complementary controls; no substitutes for the others in an agentic system.

Limit: The exact control profile remains environment-specific and must be threat-modeled per product. This is a research-synthesis design input; it does not prove product adoption, operational maturity, independent attestation or outcome.

  • OpenAI — OpenAI, Apache-2.0 repository; analysis-only excerpts
  • OWASP Foundation — OWASP Foundation, CC-BY-SA-4.0 standard; analysis-only excerpts
  • OpenSSF SLSA — OpenSSF SLSA, CC-BY-4.0 specification; analysis-only excerpts