AI security research / cases

Case files

Investigations, evidence and outcomes.

900 Cases · Page 45 of 45

Search & filter

On this page only.

Published cases

Case

Barracuda intrusion used retained vendor cloud access and ARTEX-pattern automation

Barracuda reported that an attacker used a former AI vendor's still-active cloud access key after the vendor's development environment had been compromised. The investigation described ARTEX-pattern automated activity across cloud resources, development services and CI workflows. The first-party response does not prove every command was selected autonomously, that all possible XDR customer integration credentials were downloaded, or that Barracuda and other

Setting: productionExploitation: confirmed in the wildEvidence: Grade BRemediation: mitigation available
incidentWITH AI4 claims
Case

Chainstack compromise crossed an unauthenticated Kubernetes MCP control plane

Chainstack described an intrusion that started from an exposed development Grafana instance with default credentials. Its data-source proxy could reach an internal Kubernetes MCP server without independent authentication, and commands there ran with cluster-administrator authority. The public incident account does not prove every action was model-chosen, that every accessible secret was exfiltrated or that the ARTEX tool lineage identifies the same operator a

Setting: productionExploitation: confirmed in the wildEvidence: Grade BRemediation: mitigation available
incidentWITH AI4 claims
Case

Codebook watermarks recover short-text identifiers under selected edits

A cooperative language model encodes short identifiers through selected output-token bias, but the extraction rate falls under some editing and instruction-tuned variants.

Setting: evaluationExploitation: not establishedEvidence: Grade CRemediation: not applicable
researchFirst seen Nov 12, 2026WITH AI4 claims
Case

MetaPI shows low injection success can coincide with a defense that cannot finish tasks

MetaPI converts realistic agent utility tasks into adaptive prompt-injection evaluations, reusing AppWorld, VitaBench and Terminal-Bench environments. Its tests exposed cases where a defense appeared to block attack objectives while legitimate task completion also collapsed. This is a benchmark with controlled attacks and varying models. It does not establish a universal ranking of every production agent defense or that every requested harmful call was

Setting: evaluationExploitation: demonstratedEvidence: Grade BRemediation: not applicable
researchAGAINST AI3 claims
Case

MLflow FileStore model_id permitted metadata writes outside the run directory

An MLflow tracking API used an untrusted model identifier as a filesystem path component in FileStore, allowing out-of-run metadata writes if the service was reachable.

Setting: research demonstrationExploitation: demonstratedEvidence: Grade BRemediation: patch available
vulnerabilityFirst seen Oct 26, 2026AGAINST AI4 claims
Case

PatchHolmes improves fixing-commit ranking without proving patch correctness

A local listwise coding agent inspects retrieved commit candidates to link CVEs to likely fixing changes, but selection accuracy does not establish that a patch is safe or complete.

Setting: evaluationExploitation: not establishedEvidence: Grade BRemediation: not applicable
researchFirst seen Nov 1, 2026WITH AI4 claims
Case

pgEdge MCP row counting permitted subqueries outside the requested table

The pgEdge Postgres MCP Server count_rows tool accepted a where expression containing a subquery against an unrelated table. Repeated counts could reveal facts about that table even when the MCP tool did not expose it by name. The upstream reproducer does not establish a production data leak or a bypass of native PostgreSQL SELECT restrictions.

Setting: research demonstrationExploitation: demonstratedEvidence: Grade BRemediation: patch available
vulnerabilityAGAINST AI3 claims
Case

Trusted agent intent profiles improve scope judgments at a measurable utility cost

Microsoft researchers test task-purpose profiles as a defense against agent goal drift and injected instructions, without demonstrating deterministic prevention of unauthorized effects.

Setting: simulationExploitation: not establishedEvidence: Grade CRemediation: not applicable
researchFirst seen Nov 12, 2026AGAINST AI4 claims