AgentShield: Deception-Based Compromise Detection for Tool-Using LLM Agents

Defenses against indirect prompt injection IPI in tool-using LLM agents share two structural weaknesses. First, they all attempt to prevent attacks rather than detect the compromises that slip through. Second, they have only been evaluated in English, leaving users of low-resource languages such ...

MonitoringBench: Semi-Automated Red-Teaming for Agent Monitoring

We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation,...

Oracle Poisoning: Corrupting Knowledge Graphs to Weaponise AI Agent Reasoning

We define Oracle Poisoning, an attack class in which an adversary corrupts a structured knowledge graph that AI agents query at runtime via tool-use protocols, causing incorrect conclusions through correct reasoning. Unlike prompt injection, Oracle Poisoning manipulates the data agents reason ove...

CVE-2026-42576 apko `DiscoverKeys` has a panic on non-rsa jwks key that causes crash during key discovery

apko allows users to build and publish OCI container images built from apk packages. Prior to version 1.2.7, DiscoverKeys in pkg/apk/apk/implementation.go unconditionally type-asserts JWKS keys as rsa.PublicKey without checking the key type. If a repository JWKS endpoint returns a non-RSA key e.g...

CVE-2026-42575

apko allows users to build and publish OCI container images built from apk packages. Prior to version 1.2.7, apko verifies the signature on APKINDEX.tar.gz but never compares individually downloaded .apk packages against the checksum recorded in the signed index. The checksum is parsed and...

CVE-2026-42575

CVE-2026-42575 affects chainguard/apko: before v1.2.7, apko verifies APKINDEX.signed index but does not compare individually downloaded .apk checksums to the index checksum. The ChecksumString() is parsed but never cross-checked with the downloaded package’s control hash in getPackageImpl(), allo...

OESA-2026-2283 sed security update

Sed is a non-interactive command-line text editor. A stream editor is used to per-form basic text transformations on an input stream a file or input from a pipeline. Security Fixes: When sed is invoked with both -i in-place edit and --follow-symlinks, the function opennextfile performs two...

OESA-2026-2281 sed security update

Sed is a non-interactive command-line text editor. A stream editor is used to per-form basic text transformations on an input stream a file or input from a pipeline. Security Fixes: When sed is invoked with both -i in-place edit and --follow-symlinks, the function opennextfile performs two...

CVE-2026-42301

pyp2spec generates working Fedora RPM spec file for Python projects. Prior to version 0.14.1, pyp2spec was writing PyPI package metadata e.g. the summary field into the generated spec file without escaping RPM macro directives. When a packager then runs rpmbuild, those directives get evaluated, s...

EUVD-2026-28884

bubblewrap is a low-level unprivileged sandboxing tool. From version 0.11.0 to before version 0.11.2, if bubblewrap is installed in setuid mode then the user can use ptrace to attach to bubblewrap and control the unprivileged part of the sandbox setup phase. This allows the attacker to arbitraril...

SUSE CVE-2026-39817

The "go tool pack" subcommand usually used only by the compiler as an internal tool with known-good inputs does not sanitize output filenames. Extracting a malicious archive file with the "pack" subcommand can write files to arbitrary locations on the filesystem...

CVE-2026-39820 vulnerabilities

Vulnerabilities for packages: prometheus, kubernetes, harbor, juicefs, thanos, opencost, telegraf, cilium, kubernetes-csi-node-driver-registrar, grafana-pyroscope, coredns, undock, nerdctl, kubernetes-dashboard, act, grafana-operator, kots, kyverno, datadog-agent, prometheus-adapter, pgtimetable,...

Linux Distros Unpatched Vulnerability : CVE-2026-41493

The Linux/Unix host has one or more packages installed that are impacted by a vulnerability without a vendor supplied patch available. - YARD is a Ruby Documentation tool. Prior to version 0.9.42, a path traversal vulnerability was discovered in YARD when using yard server to serve documentation...

CVE-2026-44284

FastGPT is an AI Agent building platform. Prior to version 4.14.17, FastGPT had an inconsistent SSRF protection gap in MCP tool URL handling. The direct MCP preview/run endpoints already rejected internal/private network URLs, but the MCP tool create/update endpoints could still save an internal...

CVE-2026-42346 Postiz: TOCTOU DNS rebinding bypasses all SSRF URL validation paths

Postiz is an AI social media scheduling tool. From version 2.16.6 to before version 2.21.7, all SSRF protections added in v2.21.4–v2.21.6 share a fundamental TOCTOU Time-of-Check-Time-of-Use vulnerability: isSafePublicHttpsUrl resolves DNS to validate the target IP, but subsequent fetch calls...

CVE-2026-44284

FastGPT is an AI Agent building platform. Prior to version 4.14.17, FastGPT had an inconsistent SSRF protection gap in MCP tool URL handling. The direct MCP preview/run endpoints already rejected internal/private network URLs, but the MCP tool create/update endpoints could still save an internal...

CVE-2026-44284 FastGPT: Stored MCP tool URL SSRF in FastGPT workflow execution

FastGPT is an AI Agent building platform. Prior to version 4.14.17, FastGPT had an inconsistent SSRF protection gap in MCP tool URL handling. The direct MCP preview/run endpoints already rejected internal/private network URLs, but the MCP tool create/update endpoints could still save an internal...

CVE-2026-44284

FastGPT (before 4.14.17) had an SSRF flaw in MCP tool URL handling. An authenticated user with permission to create/manage MCP toolsets could store an internal endpoint (e.g., http://localhost:3000/mcp) via the MCP tool create/update endpoints. The stored URL could be used later by the backend wo...

EUVD-2026-28867

FastGPT is an AI Agent building platform. Prior to version 4.14.17, FastGPT had an inconsistent SSRF protection gap in MCP tool URL handling. The direct MCP preview/run endpoints already rejected internal/private network URLs, but the MCP tool create/update endpoints could still save an internal...

CVE-2026-44284 FastGPT: Stored MCP tool URL SSRF in FastGPT workflow execution

FastGPT is an AI Agent building platform. Prior to version 4.14.17, FastGPT had an inconsistent SSRF protection gap in MCP tool URL handling. The direct MCP preview/run endpoints already rejected internal/private network URLs, but the MCP tool create/update endpoints could still save an internal...