< ciso
brief />
CISA Flags Exploited Flaws as AI Agent Security Gaps Surface

CISA Flags Exploited Flaws as AI Agent Security Gaps Surface

Coverage: 05 Aug 2026 (UTC)

< view all daily briefs >

Three vulnerabilities in Langflow, Apache Tomcat and N-able N-central joined CISA's Known Exploited Vulnerabilities catalog on August 5, 2026, and federal agencies have until August 7 to apply fixes. Critical disclosures in Ruby on Rails, Gitea and the Paperclip AI agent control plane add to the patching workload. AWS, Palo Alto Networks and Cloudflare detailed new controls for governing AI agents and developer workflows. Identity, certification and cloud infrastructure updates complete the day.

Actively Exploited and Critical Flaws Demand Patching

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) added three vulnerabilities to its KEV catalog on August 5, 2026, citing evidence of active exploitation. The most severe, CVE-2026-9198 (CVSS 9.8), is a code injection flaw in Langflow that allows unauthenticated remote code execution on default deployments. It was fixed in version 1.10.1 in July 2026. CVE-2026-34486 (CVSS 7.5) affects Apache Tomcat and allows the encryption of sensitive cluster messages to be bypassed. Fixes were released in April 2026 in Tomcat versions 11.0.21, 10.1.54 and 9.0.117. The additions also include CVE-2026-18556 (CVSS 8.2), an authentication bypass in N-able N-central. CVE-2026-18577, an incomplete prior fix that prompted a new patch, is listed alongside it.

Threat intelligence ties exploitation of the Tomcat flaw to an AI-enabled, Chinese-speaking actor tracked under the aliases knaithe and KnYuan. The actor used the DeepSeek capability within the Hermes Agent framework to autonomously identify and exploit internet-facing targets. It paired AI-driven reconnaissance with manual follow-up operations against multiple products and attempted exploitation across hundreds of targets. It also narrowed its scope to conserve AI compute. No public details are available yet on how the Langflow vulnerability is being weaponized, although open-source AI development platforms have been repeatedly targeted in recent months. Federal Civilian Executive Branch agencies have been directed to apply the available fixes by August 7, 2026.

A separate critical issue in Ruby on Rails, CVE-2026-66066 (9.5/10), was disclosed on July 30 and affects the Active Storage component. The flaw has been nicknamed “KindaRails2Shell.” It abuses unfuzzed libvips image operations that Active Storage leaves enabled by default. Crafted uploads can read arbitrary files or escalate to remote code execution, potentially without authentication. Exposed data can include environment variables, Rails secrets, database credentials, cloud keys and the secret_key_base. Fixed Rails releases are 7.2.3.2, 8.0.5.1 and 8.1.3.1, and libvips must also be updated to version 8.13 or later. Proof-of-concept code has been published, so the guidance is to treat the issue as a potential secret-exposure incident rather than a routine patch. Recommended steps are to:

Patch the framework and libvips, audit all image-upload features, and search logs and object stores for suspicious uploads. Then rotate secret_key_base and other credentials, invalidate sessions, and investigate downstream services for credential exposure. Longer-term measures include isolating image processing in sandboxed, minimal-privilege workers and applying strict allowlists to uploads.

Gitea disclosed CVE-2026-59774 (CVSS 9.8), a file-read vulnerability affecting versions 1.22.1 through 1.27.0. Unauthenticated attackers could submit crafted Org-mode markup to the POST /{owner}/{repo}/markup endpoint and read any file accessible to the Gitea service account. The root cause was Gitea initializing the go-org renderer with the library's default ReadFile callback. That callback let the #+INCLUDE directive return absolute filesystem paths. The bug does not directly execute code, but it can be escalated. Reading app.ini can expose INTERNAL_TOKEN, which an attacker may use to inject a Git hook via the internal logger and then trigger during an anonymous clone. Gitea 1.27.1 fixes the issue by overriding ReadFile so include paths render as plain content. Cloud instances will be auto-upgraded during maintenance. Self-hosted administrators whose markup endpoint was reached on a vulnerable build are advised to rotate the internal token, OAuth and JWT signing material, and database credentials. They should also review repository hook directories for unexpected executables. Gitea reported no observed exploitation in the wild as of August 5, 2026.

AI Agent Platforms Face Trust and Governance Tests

Oasis Security disclosed multiple Paperclip vulnerabilities in the open-source AI agent control plane. The flaws allow attackers to execute arbitrary commands by importing and starting a malicious agent.

The most severe, CVE-2026-41679 (CVSS 10.0), affects network-accessible authenticated deployments that use Paperclip's default registration configuration. An attacker can register, authorize a CLI session and import a crafted .paperclip.yaml file. The import creates a company and an agent that runs attacker-supplied commands with the server process's privileges.

A second path, GHSA-x8hx-rhr2-9rf7 (CVSS 9.6), abuses the default local_trusted mode with DNS rebinding. Attacker JavaScript loaded in a browser can later install and activate an agent on a local instance without credentials.

A third issue, GHSA-xfqj-r5qw-8g4j (CVSS 8.3), left several API routes without consistent authentication or company access checks. Those routes exposed issue data, agent documentation and health information.

The source tagged v2026.416.0 addresses the flaws. It requires instance-administrator rights for imports that create companies, adds company access checks and enables a private-hostname guard against rebound hostnames. It also tightens route-level authentication and reduces unauthenticated health responses. Rapid7 has published a Metasploit module automating the CVE-2026-41679 attack chain, and NVD/CISA classification notes proof-of-concept exploitation. No in-the-wild exploitation had been confirmed as of August 5, 2026. Operators are advised to upgrade to v2026.416.0 or later and review registration and deployment exposure settings.

Cloudflare CIO Sam Rhea described Cloudflare OS, an internal platform built to let employees deploy AI agents safely while protecting systems and data. The platform runs ephemeral cloud-based workspaces accessed through Zero Trust authentication. Data access is governed by Model Context Protocol (MCP) servers that scope agent permissions to a user's existing rights in each system of record. Cloudflare typically implements its own MCP servers to add rate limits and regional controls. All model inference traffic is routed through an AI Gateway that enables logging, DLP filtering, auditing, and model- or role-based gating. For engineering work, the Cloudflare Engineering Codex acts as a context layer. Agents consult it when reviewing merge requests, designs and incident reports, with the aim of preventing risky changes.

Palo Alto Networks introduced a direct integration between Prisma AIRS and Claude Enterprise. Incoming requests are cryptographically verified using encrypted edge key-value maps. Existing Enterprise DLP policies then classify prompts in real time and return synchronous allow or deny verdicts before Claude processes the input. Prisma AIRS also inspects payloads for LLM-specific risks such as prompt injection. It provides deny responses, optional user-facing explanations and unique audit correlation codes that feed security telemetry.

AWS announced a partnership with Anthropic and OpenAI to integrate AWS Continuum into coding assistants such as Claude Code, OpenAI Codex and Kiro. The Preview of Continuum for code vulnerabilities combines multiple models through an orchestration layer, or harness. The harness selects a model for each task, connects to customer environments and enforces guardrails. Continuum scans code on demand and prioritizes findings using context such as IAM, network topology, exposure surfaces and configurations. It then validates issues in a sandbox and returns prioritized results to the coding assistant. AWS is also working with the Frontier Model Forum and model providers on shared safety standards and security performance benchmarking.

Identity Controls and Industrial Assurance

AWS IAM Identity Center now lets organizations disable automatic management of AWS account access when creating a new instance. Administrators can limit Identity Center to managing AWS applications. This avoids creating its service-linked role in member accounts and reduces the access surface. The option is presented during initial configuration and is not retroactive for existing instances. Account management can be enabled later through instance settings or the UpdateInstance API.

Amazon Cognito added provisioned limits for user pools, available from July 6, 2026. The feature separates the account-level maximum, set through AWS Service Quotas, from a provisioned limit managed in the Cognito console or via the UpdateProvisionedLimit API. The change reduces support-ticket lead times of 10–14 days to minutes. Customers are billed only for provisioned capacity above default limits, in increments as small as 1 RPS for one day.

Fortinet achieved IEC 62443-4-2 Security Level 4 certification for FortiOS v7.6.x. SL4 is the highest technical assurance level for industrial automation and control system components. The certification builds on Fortinet's earlier IEC 62443-4-1 Maturity Level 2 recognition for its Secure Product Development Lifecycle. It applies across FortiGate, FortiGate Rugged, FortiWiFi and FortiGate VM platforms running FortiOS v7.6.x. The assessment covered all seven foundational requirement categories of IEC 62443-4-2, including identification and authentication, system integrity, data confidentiality and restricted data flows. The standard is referenced in regulatory and procurement requirements for operational technology in sectors such as utilities, transportation, manufacturing and energy.

Cloud Platform Updates for Data and AI Workloads

Amazon DynamoDB gained generally available native vector search, with real-time indexing and approximate nearest neighbor queries. AWS cites single-digit millisecond latency and 99%+ recall as datasets grow to billions or trillions of vectors. The feature supports attribute filtering and pay-for-what-you-use pricing. Amazon Aurora Serverless can now reach up to 12 ACUs within a single second during scale-up events and continue scaling to 256 ACUs. The enhancement is enabled by default for platform versions 3 and 4, and clusters on versions 1 and 2 can upgrade directly to version 4.

AWS Lambda functions running outside a VPC with 2 GB of memory or more can now use scalable network bandwidth. Bandwidth grows with memory, from 625 Mbps at 2 GB to 3,000 Mbps at 10 GB. Activation requires a request through AWS Service Quotas. The capability carries no additional charge and is available in all commercial Regions. AWS Glue Data Quality removed the extra charge for anomaly detection in Glue ETL jobs. It also added an observation mode that uses a constant baseline instead of linear trend extrapolation, which reduces false positives when data arrives at irregular intervals.

The AWS Glue Data Catalog preview now supports metadata exports to S3 Tables. Technical and business metadata are written to the managed aws-catalog S3 table bucket in Apache Iceberg format, enabling SQL audits and time travel queries. Semantic search in the preview now extends to catalogs encrypted with customer managed AWS KMS keys. The feature is initially available in US East (N. Virginia), US East (Ohio), US West (Oregon) and Europe (Ireland), and S3 Tables usage is charged at standard S3 pricing.

Google Cloud moved Filestore onto Colossus, its foundational distributed storage system. The change decouples storage capacity from IOPS and supports GKE multishares starting at 10 GiB. Access controls integrate with Google Cloud IAM, NFS UIDs/GIDs and IP ACLs.