⚡ Weekly Recap: AI Goes Rogue, Metabase 0-Day, MCP Supply

Flowise's Custom MCPs feature launches local MCP servers by executing operating-system commands (for example via npx), but Flowise has a minimal authentication and authorization model with no role-based access controls, and versions before 3.0.1 run without authentication by default unless it is explicitly configured. As a result, an unauthenticated attacker with network access to an affected instance can invoke this functionality to run unsandboxed OS commands. Successful exploitation yields full unauthenticated remote code execution on the host with the privileges of the Flowise service, with high impact to confidentiality, integrity, and availability (CVSS 3.1: 9.8). Any Flowise deployment prior to 3.0.1 is affected, with default, unauthenticated installations reachable from untrusted networks at greatest risk. JFrog has published technical details and a PoC, EPSS assigns a 66.3% probability of exploitation within 30 days (99th percentile), and news coverage reports active exploitation with 12,000+ instances exposed on the internet.

Do: Upgrade Flowise to version 3.0.1 or later. Until upgraded, restrict network access to the instance (firewall rules or an authenticating reverse proxy) and explicitly enable authentication, since default installations before 3.0.1 run without it. Review internet-exposed instances for signs of compromise, such as unexpected npx launches or child OS processes spawned by the Flowise service.

The terraform-mcp-server before version 1.1.0 is vulnerable to a server-side request forgery issue in the streamable-HTTP transport that may allow an unauthenticated remote client to redirect the server's Terraform API requests, and the server-side authorization token, to an attacker-controlled endpoint. This vulnerability, CVE-2026-14869, is fixed in terraform-mcp-server 1.1.0.

An issue was discovered in Django 5.2 before 5.2.17 and 6.0 before 6.0.8. GeoDjango spatial lookups optimistically parse the right-hand-side value as a raster by passing it to the `django.contrib.gis.gdal.GDALRaster` constructor. Any value used in a spatial lookup against a `GeometryField` or `RasterField` reaches this constructor, including untrusted input, for example a spatial-field filter submitted through the Django admin changelist query string by a staff user with view permission. A `dict`, or a `str` holding its JSON representation, is opened in write mode regardless of the constructor's `write=False` default, allowing a file with an attacker-chosen name and contents to be written through a file-backed GDAL driver. Any other `str` is treated as a datasource, allowing an outbound network request through a GDAL virtual filesystem handler. Writing a file to a location later imported by the application can result in remote code execution. Earlier, unsupported Django series (such as 5.1.x, 5.0.x, and 4.2.x) were not evaluated and may also be affected. Django would like to thank Bence Nagy, localhost-detect, and kimchunbok_ for reporting this issue.

The terraform-mcp-server before version 1.1.0 is vulnerable to an authorization bypass in the streamable-HTTP stateful transport mode that may allow a user who obtains another user's MCP session ID to have their tool calls executed using that user's Terraform credentials. This vulnerability, CVE-2026-16496, is fixed in terraform-mcp-server 1.1.0.

The terraform-mcp-server before version 1.1.0 is vulnerable to a cross-tenant credential reuse issue in the streamable-HTTP stateless transport mode that may allow one user's Terraform token to be used to execute tool calls on behalf of subsequent users. This vulnerability, CVE-2026-16498, is fixed in terraform-mcp-server 1.1.0.

The affected Thermo Fisher Applied Biosystems Genetic Analyzers are vulnerable because .fsa/.hid output files can be edited. An attacker could tamper with these files, altering DNA data and resulting in inaccurate DNA test outcomes.

A vulnerability in the Agent Development Kit (ADK) allows for continuation forgery in tool confirmations. An attacker who is able to manipulate or inject events into the session history can execute unauthorized tools by forging a tool confirmation response. This is possible because the framework did not verify if the target tool was registered to the executing agent, did not validate if the tool actually required confirmation, and did not match the confirmation arguments against the original tool call event in the history.

A heap-buffer-overflow vulnerability exists in the nothings stb TrueType library, up to version 1.26, that is used for parsing TrueType font files. The vulnerability exists in the stbtt__GetGlyphShapeTT() function within the nothings stb_truetype.h library when parsing malformed TTF (TrueType Font) files. The vulnerability resides in the glyph data parsing path. An attacker can craft a malformed TTF file with an inflated endPtsOfContours value and truncate the remaining glyph data. When an application utilizing stb_truetype.h (such as various game engines or graphics software) attempts to load, bake, or render this malformed font via stbtt_GetGlyphShape(), the parser will attempt to read past the end of the glyph data buffer, triggering the out-of-bounds read.

N-able's N-central RMM platform contains an authentication bypass (CWE-288) in which an alternate path or channel allows requests to skip the normal authentication check. An unauthenticated attacker can trigger it by sending requests through that alternate path without valid credentials, gaining unauthorized access to the N-central management interface; because N-central is remote monitoring and management software run by managed service providers, such access can expose management functions across downstream customer environments. N-able N-central deployments are affected; the available data does not specify affected version ranges or fixed builds. The flaw was added to CISA's Known Exploited Vulnerabilities catalog on 2026-08-04, confirming exploitation in the wild, with a high EPSS of 40.2% (99th percentile), no CVSS score yet, no public PoC, and undetermined ransomware use; related reporting describes a recent string of N-central hotfixes, including fixes for unauthenticated remote code execution.

Do: Apply N-able's hotfix for N-central immediately per the vendor's instructions, since the flaw is in CISA KEV and BOD 26-04 applies (for cloud-hosted N-central, apply mitigations per BOD 26-04 or discontinue use if mitigations are unavailable). Until patched, restrict internet exposure of the N-central server to trusted management networks and review authentication logs for unexpected access, following CISA's forensics triage guidance. Check the vendor advisory for the exact fixed build, as the available data does not list affected or fixed version numbers.

CVE-2026-18577 is an authentication bypass (CWE-288) in N-able's N-central RMM platform caused by an incomplete patch for the earlier vulnerability CVE-2026-18556. Because the original fix can be bypassed via an alternate path or channel, a remote, unauthenticated attacker needs no privileges or user interaction, though the attack requires meeting exploit-specific conditions (high attack complexity, CVSS 4.0: 8.2 High). Successful exploitation lets the attacker bypass authentication and take over N-central accounts, gaining high-impact access to the management console with limited direct effects on downstream services. All N-central versions through 2026.3.1 are affected. Exploitation is confirmed in the wild: the flaw was added to CISA's KEV on 2026-08-03, EPSS estimates a 54.1% chance of exploitation within 30 days (99th percentile), and news reports say attackers kept exploiting it even as N-able shipped successive hotfixes; ransomware use is currently unknown.

Do: Upgrade N-central to a fixed release or hotfix newer than 2026.3.1 following N-able's advisory — the vendor has issued multiple successive hotfixes for this and related N-central flaws, so verify you are on the latest build. Because the flaw is on CISA's KEV, federal and BOD 26-04-bound organizations must apply vendor mitigations promptly or discontinue use of the product if patching is unavailable. Limit internet exposure of N-central portals and review accounts for signs of takeover or unauthorized access.

Insufficient input validation in Amazon Bedrock AgentCore harness might allow an authenticated remote user to execute configured tools bypassing model invocation and security controls via crafted content blocks in conversation messages. AWS has addressed this issue. No customer action is required.

Use after free in Aura in Google Chrome on Linux prior to 151.0.7922.109 allowed a remote attacker to potentially perform a sandbox escape via a crafted HTML page. (Chromium security severity: Critical)

As part of Cisco's ongoing commitment to proactive security and product quality, the Cisco IOS XE Software engineering team has conducted a comprehensive internal security review. This review resulted in software hardening releases that address multiple internally discovered vulnerabilities. The vulnerabilities tracked by CVE-2026-20272 are related to issues with improper neutralization of special elements that are grouped under the Common Weakness Enumeration (CWE) Pillar CWE-74.

As part of Cisco's ongoing commitment to proactive security and product quality, the Cisco Catalyst SD-WAN engineering team has conducted a comprehensive internal security review. This review resulted in software hardening releases that address multiple internally discovered vulnerabilities. The vulnerabilities tracked by CVE-2026-20303 are related to improper input validation issues that are grouped under the Common Weakness Enumeration (CWE) CWE-20.

Protection mechanism failure in Windows Event Logging Service allows an authorized attacker to disclose information over a network.

Paperclip is a Node.js server and React UI that orchestrates a team of AI agents to run a business. Prior to version 2026.416.0, an unauthenticated attacker can achieve full remote code execution on any network-accessible Paperclip instance running in `authenticated` mode with default configuration. No user interaction, no credentials, just the target's address. The chain consists of six API calls. The attack is fully automated, requires no user interaction, and works against the default deployment configuration. Version 2026.416.0 patches the issue.

Modification of assumed-immutable data (maid) in Azure Active Directory allows an authorized attacker to elevate privileges over a network.

Deserialization of untrusted data in Azure Service Bus allows an authorized attacker to execute code over a network.

Improper authentication in Azure SQL Database allows an unauthorized attacker to elevate privileges over a network.

Origin validation error in Windows Network Address Translation (NAT) allows an unauthorized attacker to perform spoofing over an adjacent network.

Improper preservation of SQL mode when renaming databases in cPanel allows execution of SQL in root context.

A vulnerability in Veeam Service Provider Console allowing arbitrary file write on the management server, which can lead to remotecode execution.

A vulnerability in Veeam Service Provider Console allowing an unauthenticated attacker to impersonate a managed agent andobtain that agent's credentials.

'.../...//' in Microsoft Entra Provisioning Service (SyncFabric) allows an authorized attacker to elevate privileges over a network.

Missing authorization in Azure SRE Agent allows an authorized attacker to elevate privileges over a network.

Missing authentication for critical function in Microsoft Planetary Computer Pro allows an unauthorized attacker to elevate privileges over a network.

In the Linux kernel, the following vulnerability has been resolved: netfilter: conntrack: tcp: do not force CLOSE on invalid-seq RST without direction check An unintended behavior in the TCP conntrack state machine allows a connection to be forced into the CLOSE state using an RST packet with an invalid sequence number. Specifically, after a SYN packet is observed, an RST with an invalid SEQ can transition the conntrack entry to TCP_CONNTRACK_CLOSE, regardless of whether the RST corresponds to the expected reply direction. The relevant code path assumes the RST is a response to an outgoing SYN, but does not validate packet direction or ensure that a matching SYN was actually sent in the opposite direction. As a result, a crafted packet sequence consisting of a SYN followed by an invalid-sequence RST can prematurely terminate an active NAT entry. This makes connection teardown easier than intended. So, tighten the state transition logic to ensure that RST-triggered CLOSE transitions only occur when the RST is a valid response to a previously observed SYN in the correct direction.

In the Linux kernel, the following vulnerability has been resolved: net: openvswitch: reject oversized nested action attrs Open vSwitch stores generated flow actions as nlattrs, whose nla_len field is u16. Commit a1e64addf3ff ("net: openvswitch: remove misbehaving actions length check") allowed the total sw_flow_actions stream to grow beyond 64 KiB, which is valid, but also removed the last guard preventing a generated nested action attribute from exceeding U16_MAX. An oversized generated container can thus be closed with a truncated nla_len. A later dump or teardown then walks a structurally different stream than the one that was validated. In particular, an oversized nested CLONE/CT action may cause subsequent bytes in the generated stream to be interpreted as independent actions. Keep the larger total-action-stream behavior, but make nested action close reject generated containers that do not fit in nla_len, and return the error through all callers. For recursive SAMPLE, CLONE, DEC_TTL, and CHECK_PKT_LEN builders, trim resource-owning action-list tails in reverse construction order before discarding failed wrappers, so resources copied into the rejected tails are released before the wrappers are removed. Most failed outer wrappers are discarded by truncating actions_len after child resources have been released. CHECK_PKT_LEN also trims its parent after branch resources are gone. SET/TUNNEL close failures unwind their known tun_dst ownership directly, and SET_TO_MASKED has no external ownership and truncates on close failure.

In the Linux kernel, the following vulnerability has been resolved: KVM: x86: Check for invalid/obsolete root *after* making MMU pages available Check for a "stale" page fault, i.e. for an invalid and/or obsolete root, after making MMU pages available for the shadow MMU. If reclaiming shadow pages zaps an in-use root, i.e. marks it invalid, then KVM will attempt to map memory into an invalid root. On its own, populating an invalid root is "fine", but because child shadow pages inherit their parent's role, any children created during the map/fetch will be created as invalid pages, thus violating KVM's invariant that invalid pages are never on the list of active MMU pages. Note, the underlying flaw has existed since KVM first started tracking invalid roots in 2008 (commit 2e53d63acba7, "KVM: MMU: ignore zapped root pagetables"), but the true badness only came along in 2020 (Linux 5.9) with the invariant that invalid shadow pages can't be on the list of active pages. Note #2, inheriting role.invalid when creating child shadow pages is also far from ideal; that flaw will be addressed separately.

In the Linux kernel, the following vulnerability has been resolved: sctp: don't free the ASCONF's own transport in DEL-IP processing sctp_process_asconf() caches the transport the ASCONF chunk is processed against in asconf->transport (== chunk->transport, set once in sctp_rcv()). For an ASCONF located through its Address Parameter by __sctp_rcv_asconf_lookup(), that cached transport corresponds to the Address Parameter, which need not be the packet's source address. sctp_process_asconf_param() rejects a DEL-IP for the packet source address (ADDIP D8, SCTP_ERROR_DEL_SRC_IP), but nothing protects asconf->transport. A single ASCONF can therefore carry, in order: [Address Parameter L] [DEL-IP L] [DEL-IP 0.0.0.0] where L differs from the source. The DEL-IP for L passes the D8 check and calls sctp_assoc_rm_peer() on the transport that asconf->transport still points at, freeing it (RCU-deferred). The following wildcard DEL-IP then reuses the now-dangling asconf->transport in sctp_assoc_set_primary() and sctp_assoc_del_nonprimary_peers(): set_primary() dereferences the freed transport (->ipaddr, ->state) and plants the dangling pointer into asoc->peer.primary_path / active_path, and del_nonprimary_peers(), keeping only the pointer that is no longer on the list, removes every real transport, leaving the association with a transport_count of 0 and primary_path/active_path pointing at freed memory. Reject a DEL-IP that targets the transport the ASCONF is being processed against, mirroring the existing source-address guard, so the wildcard branch can never reuse a freed transport.

WordPress core contains a reflected cross-site scripting (CWE-79) vulnerability on the login screen that can be triggered without authentication. An attacker must socially engineer a victim into clicking a crafted link or interacting with an attacker-hosted third-party website, so explicit user interaction is required for exploitation. If successful, the attacker can run malicious script in the victim's browser on the targeted WordPress site, and — under conditions outside the attacker's direct control — escalate the XSS to remote code execution (PHP code execution). Every version of WordPress is affected; version 7.0.3 fixes the flaw and the fix has been backported to all branches back to 4.7. No public proof-of-concept is known and the issue is not yet in CISA's KEV catalog, though EPSS assigns a 31.2% probability of exploitation within 30 days (98th percentile), indicating elevated near-term risk.

Do: Update to WordPress 7.0.3, or apply the backported security update on your branch (backports are available for all branches back to 4.7). Until patched, exercise caution with unsolicited login-page links and consider monitoring authentication and admin-action logs for suspicious activity, since exploitation depends on user interaction. Because the RCE escalation depends on environmental conditions, treat all unpatched WordPress sites as exposed even if no in-the-wild exploitation is yet confirmed.

The `@ai-sdk/harness-opencode` tool is an HarnessV1 adapter backed by @openai/codex-sdk, which drives the codex command line interface. Prior to version 1.0.29, the tool relay authorizes requests from any process whose command line contains an allowed helper script path (the Codex CLI shim). This allows untrusted code executing in the sandbox to invoke arbitrary host-exposed tools, including secret lookups, deployment operations, and cloud API calls without a corresponding model-authorized tool-call event. Exploitation requires a Linux environment (the vulnerable fallback checks `process.platform === 'linux'` and reads `/proc`); an active harness session with one or more host-provided tools; and untrusted code executing in the sandbox (e.g. a malicious dependency, build script, or lifecycle hook) The fix in version 1.0.29 removes the process-path authorization fallback entirely. Relay requests are now only accepted after exact, short-lived, one-time authorization matching the tool name and input from a bridge-observed model event. Some workarounds are available. Do not run the Codex harness on untrusted repositories or with untrusted dependencies, and/or limit host-exposed tools to non-sensitive operations when working with untrusted code.

The `@ai-sdk/harness-opencode` tool connects HarnessAgent to OpenCode through a sandboxed bridge. Prior to version 1.0.28, the tool relay authorizes requests from any process whose command line contains an allowed helper script path (`host-tool-mcp.mjs`). This allows untrusted code executing in the sandbox to invoke arbitrary host-exposed tools including secret lookups, deployment operations, and cloud API calls without a corresponding model-authorized tool-call event. Exploitation requires a Linux environment (the vulnerable fallback checks `process.platform === 'linux'` and reads `/proc`); an active harness session with one or more host-provided tools; and untrusted code executing in the sandbox (e.g. a malicious dependency, build script, or lifecycle hook) The fix in version 1.0.28 removes the process-path authorization fallback entirely. Relay requests are now only accepted after exact, short-lived, one-time authorization matching the tool name and input from a bridge-observed model event. Some workarounds are available. Do not run the OpenCode harness on untrusted repositories or with untrusted dependencies, and/or limit host-exposed tools to non-sensitive operations when working with untrusted code.

CVE-2026-65400 is a critical (CVSS 9.8) improper authentication flaw (CWE-287) in Apple macOS's Screen Sharing service, caused by an authentication state-management defect. An attacker who can reach a vulnerable Mac's Screen Sharing service over the network can authenticate without valid credentials, gaining full remote access with high impact to confidentiality, integrity, and availability. All three currently supported macOS branches are affected: Sequoia, Sonoma, and Tahoe, in versions prior to the fixed releases. The flaw is being actively exploited on the internet, with public reporting that attackers use the bypass to deploy Monero cryptominers, and CISA added it to the Known Exploited Vulnerabilities catalog on 2026-08-18. EPSS estimates a 9.9% probability of exploitation within 30 days (95th percentile).

Do: Upgrade to macOS Sequoia 15.7.9, macOS Sonoma 14.8.9, or macOS Tahoe 26.6.1 (or later) immediately; patching is mandatory for federal agencies under CISA BOD 26-04 given the KEV listing. As an interim mitigation, disable Screen Sharing or restrict it via firewall/VPN so VNC (port 5900) is not reachable from the internet. Review internet-exposed Macs for signs of compromise, especially unexplained Monero miner processes or abnormal CPU usage.

Missing authorization in Microsoft Teams allows an unauthorized attacker to elevate privileges over a network.

A cross-site scripting (XSS) vulnerability exists in Alinto SOGo, version 5.12.7. A maliciously crafted ICS calendar invitation files allows arbitrary JavaScript execution within the authenticated SOGo webmail session. The issue occurs because SVG content embedded in the description field of an ICS file, with an onrepeat event handler, is insufficiently sanitized before being rendered in the webmail interface. A remote attacker can execute JavaScript in the victim's browser when the malicious calendar invite is viewed. Successful exploitation may allow mailbox access, email and contact theft, session hijacking, and other actions allowed by an authenticated user.