The internet is broken right now.
45 critical CVEs without a fix in the last 48 hours. Somewhere a staging server is public. It always is. 50 more already have a fix that someone still has to apply.
-
CVE-2026-131820
A server-side request forgery vulnerability in GitHub Enterprise Server allowed an authenticated user with repository write access to send requests to internal services through Slack webhook validity checks, potentially leading to remote code execution on the appliance. Exploitation required Secret Protection with secret scanning and validity checks for partner patterns enabled. This vulnerability affected all versions prior to 3.23 and was fixed in versions 3.22.5, 3.21.10, and 3.20.12. This vulnerability was reported via the GitHub Bug Bounty program.
-
CVE-2026-129510
Improper Output Neutralization for Logs vulnerability in the XmlLayout and XmlLayoutSchemaLog4J of Apache log4net. Both layouts split a message holding the CDATA end marker into several CDATA sections, but searched for the marker with a culture-sensitive comparison. Ignorable characters or a combining mark around the marker made the search miss it, so the CDATA section stayed open. From 3.3.1 on, an XML parser then reads the following log events as part of that message; earlier versions leave the log file not well-formed. A party whose data reaches a log message could hide later log events from a consumer that parses the XML log. This issue affects Apache log4net: from 1.2.9 before 3.6.0. Users are recommended to upgrade to version 3.6.0, which fixes the issue.
-
CVE-2026-129851
SeaweedFS before 4.42 contains a server-side request forgery vulnerability in the VolumeServer.FetchAndWriteNeedle gRPC method that allows unauthenticated attackers to bypass the incomplete CVE-2026-73080 fix. Attackers can supply crafted replicas[].url entries or an azure_endpoint to make the volume server send requests with controlled tokens and bodies to internal services and metadata addresses.
-
CVE-2026-105624
The login flow in Zenitel Connect before 1.11.0 applies per-account lockout on repeated failed login attempts, which allows remote attackers to cause a denial of service (temporary lockout of targeted user accounts from the web interface, repeatable for as long as the attack continues) via repeated failed login requests for a known or guessed username. Intercom call handling is not affected.
-
CVE-2026-105625
The web interface in Zenitel Connect before 1.11.0 loads an integration iframe without sandboxing or origin restrictions, which allows remote attackers to present phishing content to authenticated users via a crafted frame source.
-
CVE-2026-109445
PLANET GS-5220-8P2T2S (hardware version 3.0) firmware versions before 3.2412b261008 contains a stack-based buffer overflow vulnerability that allows unauthenticated attackers to corrupt stack memory by submitting an oversized webstax_auth HTTP Cookie value that is copied into a fixed-size stack buffer without length validation. Attackers can send a crafted unauthenticated HTTP request with an oversized cookie to overwrite the saved return address and corrupt stack memory, leading to memory corruption and potential denial of service on the device.
-
CVE-2026-119531
The Live Composer WordPress plugin before 2.1.24 does not escape several attributes of one of its page-builder shortcodes before outputting them inside HTML tag attributes, allowing users with the Contributor role and above to store a payload that executes in the browser of anyone who views the post, including an administrator reviewing the pending submission.
-
CVE-2026-117521
The Dynific Addons for Elementor (formerly Anywhere Elementor) WordPress plugin before 1.3.1 does not escape a user-controlled profile value before outputting it in a quoted attribute in one of its Elementor widgets, allowing authenticated users with the Author role to store a payload that executes in the session of anyone, including an administrator, who views a post the attacker authored, resulting in Stored Cross-Site Scripting.
-
CVE-2026-114161
Improper limitation of a pathname to a restricted directory ('path traversal') in Azure CycleCloud allows an authorized attacker to elevate privileges locally.
-
CVE-2026-89126
SAP Web Dispatcher does not sufficiently validate certain network input processed by an internal component. An unauthenticated attacker could send specially crafted network traffic to trigger a memory corruption condition in the affected functionality. Successful exploitation could allow the attacker to execute arbitrary code and fully compromise the system, resulting in a high impact on confidentiality, integrity and availability of the system.
-
CVE-2026-76964
SAP NetWeaver Application Server ABAP does not sufficiently validate certain network input processed by an internal component. An unauthenticated attacker could send specially crafted network traffic to trigger a memory corruption condition in the affected functionality. Successful exploitation could allow the attacker to execute arbitrary code and fully compromise the system, resulting in a high impact on confidentiality, integrity and availability of the system.
-
CVE-2026-63655
tsParticles is a TypeScript library for creating highly customizable particle animations. Prior to 4.3.1, the HTTP mode of @tsparticles/mcp-server accepts unauthenticated POST /mcp requests in integrations/mcp-server/src/index.ts and accumulates every request-body chunk in a Buffer[] without a cumulative size limit before Buffer.concat and JSON.parse. A remote attacker can stream an arbitrarily large body to exhaust Node.js process memory and terminate the service, including in the official Docker configuration that exposes HTTP mode on 0.0.0.0:3000. Confidentiality and integrity are not affected. This issue is fixed in version 4.3.1.
-
CVE-2026-121956
FreeRDP before 3.32.2 contains a double free in the WinPR Kerberos SSPI retrieveSomeTgt function, which installs the krb5 default realm into a target principal that has an empty realm without taking ownership of it. Unauthenticated remote clients can send a crafted NLA/CredSSP u2u TGT request whose server-name component contains '@' to corrupt the heap and crash the server process.
-
CVE-2026-120632
CGI versions before 5.00 for Perl allow CPU exhaustion via a backtracking regex on the XForms multipart/related Content-Type. When a POST Content-Type contains multipart/related, the boundary and start parameters are extracted with an unanchored pattern that puts .+ on either side of the boundary capture, which backtracks in cubic time on a header that does not match. No request body is needed.
-
CVE-2026-119662
Songloft 2.9.0 before 2.13.6 contains a hardcoded credentials vulnerability that allows adjacent network attackers to authenticate as administrator using the fixed admin/admin pair. Attackers on the same network can log in via the /api/v1/auth/login endpoint to modify library and configuration settings and install JavaScript plugins running in QuickJS.
-
CVE-2026-109381
The Mage Eventpress plugin for WordPress is vulnerable to privilege escalation in versions up to, and including, 5.7.8 This is due to the mep_grant_shop_manager_access() function hooking user_has_cap and unconditionally granting manage_options whenever an attacker-controllable request marker matches — most notably $_GET['page'] starting with 'mep_' or 'mpwem_', which is never validated against an actually registered EventPress screen. This makes it possible for authenticated attackers, with Shop Manager-level access and above, to elevate their privileges to Administrator by appending ?page=mep_x to core admin URLs (options-general.
-
CVE-2026-116744
ntop nDPI through 6.0 contains an out-of-bounds read vulnerability in dissect_softether_host_fqdn() in softether.c that compares host_name using the consumed byte count instead of data length. Remote attackers can send crafted short UDP datagrams on already-classified SoftEther flows, causing strncmp to read beyond the packet payload buffer.
-
CVE-2026-99077
In the Linux kernel, the following vulnerability has been resolved: ipv4: udp: Create exceptions before socket matching Currently, when ICMP Fragmentation Needed and Redirect Message packets are locally delivered and quote a UDP packet, a FIB nexthop exception (FNHE) is only created if the kernel can match the UDP packet to an existing socket. This behavior allows off-path attackers to conduct a side-channel attack on the FNHE cache in order to discover the ephemeral port used by a connected UDP socket. Commit 6457378fe796 ("ipv4: use siphash instead of Jenkins in fnhe_hashfun()") and commit 67d6d681e15b ("ipv4: make exception cache less predictible") tried to mitigate such attacks by making it harder for attackers to discover hash collisions in the FNHE cache and by randomizing the number of exceptions a hash bucket can hold, respectively. Unfortunately, both of the mitigations can be bypassed. Instead, mitigate such attacks by always creating a FNHE, even before trying to find a matching socket. Do that by calling ipv4_update_pmtu() and ipv4_redirect(), the helpers used when the quoted packet did not originate from a socket. This means that guesses (right or wrong) from an off-path attacker will always result in a FNHE being created or updated in the cache that the attacker can observe. Pass an oif of 0, in a similar fashion to icmp_err(). This is also the oif used by the socket path for sockets that are not bound to a device. Note that this does not allow attackers to create FNHEs that they could not create before, as both helpers can already be reached with little to no validation. For example, by sending an ICMP error that quotes an ICMP Echo Reply or one that quotes a UDP source port that matches a wildcard socket. Also note that in the good case (matched socket) the above scheme comes at the cost of an extra route lookup, as the no socket helpers perform their own lookup before the one performed by ipv4_sk_update_pmtu() / ipv4_sk_redirect(). When the two resolve to different nexthops, it also results in two exceptions being created for the same destination IP. One in the FNHE cache of the nexthop resolved by the no socket helpers and another in the FNHE cache of the nexthop used by the socket.
-
CVE-2026-115517
Unrestricted network access between execution contexts in KNIME Business Hub before version 1.19.0 allows an authenticated remote user to access open network ports on executors in different execution contexts. This only poses a security issue if the IP address of the attacked executor is know and there are open network ports that provided access without further authentication. By default no network ports are open. However, depending on the running workflows, ports may be opened and provide an attack surface. As a workaround, network policies can be created manually.
-
CVE-2026-76527
As part of Cisco's ongoing commitment to proactive security and product quality, Cisco has conducted a comprehensive internal security review. This review resulted in a software hardening release that addresses multiple internally discovered vulnerabilities. The vulnerabilities tracked by CVE-2026-76527 are related to improper check or handling of exceptional conditions and are grouped under the Common Weakness Enumeration (CWE) CWE-703.
-
CVE-2026-131581
The PDF Poster – Let visitors read PDF’s without leaving the page plugin for WordPress is vulnerable to authorization bypass in all versions up to, and including, 2.6.5 This is due to the plugin not properly verifying that a user is authorized to perform an action. This makes it possible for authenticated attackers, with contributor-level access and above, to retrieve the titles and identifiers of other users' private, draft, pending, and scheduled pdfposter posts, and retrieve the title of any arbitrary post by ID across the entire site.
-
CVE-2026-123891
SDDC Manager contains an SQL-injection vulnerability due to improper sanitization of user-supplied input. Affected versions: 9.1.x.x (fixed in 9.1.1.0200); 9.0.x.x (fixed in 9.0.2.0100); 5.x (fixed in 5.2.5).
-
CVE-2026-108761
The UpdraftPlus: WP Backup & Migration Plugin for WordPress is vulnerable to second-order SQL Injection via the user profile description field in all Free versions up to, and including, 1.26.15 and all Premium versions up to, and including, 2.26.15. This makes it possible for authenticated attackers, with subscriber-level access and above, to store a malicious payload that is written into a database backup and executed as SQL when an administrator later restores that backup. Successful exploitation can modify sensitive WordPress data and, on sites where plugin installation is available, lead to remote code execution through the installation and activation of an attacker-controlled plugin.
-
CVE-2026-114927
Vulnerability in the PeopleSoft Enterprise PeopleTools product of Oracle PeopleSoft (component: Report Distribution). Supported versions that are affected are 8.61-8.63. Easily exploitable vulnerability allows high privileged attacker with network access via HTTP to compromise PeopleSoft Enterprise PeopleTools. Successful attacks of this vulnerability can result in takeover of PeopleSoft Enterprise PeopleTools. CVSS 3.1 Base Score 7.2 (Confidentiality, Integrity and Availability impacts). CVSS Vector: (CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:U/C:H/I:H/A:H).
-
CVE-2026-120699
CubeSandbox is an instant, concurrent, secure, and lightweight sandbox for AI agents. Prior to 0.5.1, any authenticated tenant permitted to call POST /sandboxes can supply metadata["host-mount"] with an arbitrary existing host path because the control plane performs no authorization check or allowed-prefix validation for the path. injectHostDirMounts forwards the path to prepareHostDirVolume, where Cubelet uses mount --rbind through the function to perform a recursive bind mount as root and exposes the host subtree to the tenant's MicroVM through virtio-fs, read-write by default. This allows host-root file access and modification, host code execution, access to the container runtime, and compromise of co-resident tenants without a kernel or virtual-machine-monitor flaw. This issue is fixed in version 0.5.1.
-
CVE-2026-126465
OpenEMR before 8.5.0 contains an SQL injection vulnerability in the patient portal api/onsiteactivityviews endpoint that allows any authenticated staff user to inject SQL through unquoted Id_BitwiseAnd Criteria values. Attackers can close the bitwise parenthesis and append OR or UNION SELECT clauses to read patient PHI, staff password hashes, MySQL authentication strings, and API tokens.
-
CVE-2026-126352
Addressable is an alternative implementation to the URI implementation that is part of Ruby's standard library. From 2.3.0 until 2.9.1, URI templates with several variables in one expression or several adjacent expressions generate ambiguous regular expressions. Affected forms include multi-variable expressions using operators other than plus and number sign, large plus or number-sign expressions, and adjacent single-variable expressions. Template#match or Template#extract can backtrack polynomially or exponentially on attacker-controlled URIs, causing denial of service on MRI Ruby before 3.2 and JRuby, with the large plus and number-sign form also affecting MRI Ruby 3.2 and TruffleRuby 40.0. URI parsing and Template#expand are unaffected. This issue is fixed in version 2.9.1.
-
CVE-2026-116318
SQL injection in PostgreSQL pg_dump allows an object creator to execute arbitrary SQL as the user restoring the dump, via an inherited column having a child-local not-null constraint whose name requires quoting. Within major version 18, minor versions before PostgreSQL 18.7 are affected. Versions before PostgreSQL 18 are unaffected.
-
CVE-2026-126249
A improper privilege management vulnerability in Fortinet FortiAuthenticator 8.0.0 through 8.0.3, FortiAuthenticator 6.6.0 through 6.6.10 may allow an authenticated attacker to escalate privileges.
-
CVE-2026-126175
A improper neutralization of special elements used in an sql command ('sql injection') vulnerability in Fortinet FortiAnalyzer 7.6.0 through 7.6.7, FortiAnalyzer 7.4.0 through 7.4.11, FortiAnalyzer 7.2.2 through 7.2.12, FortiAnalyzer Cloud 7.6.0 through 7.6.7, FortiAnalyzer Cloud 7.4.1 through 7.4.11, FortiAnalyzer Cloud 7.2.2 through 7.2.13, FortiManager 7.6.0 through 7.6.7, FortiManager 7.4.0 through 7.4.11, FortiManager 7.2.2 through 7.2.12, FortiManager Cloud 7.6.0 through 7.6.7, FortiManager Cloud 7.4.1 through 7.4.11, FortiManager Cloud 7.2.2 through 7.2.13 may allow an authenticated attacker to execute unauthorized code or commands via crafted JSON-RPC requests.
-
CVE-2026-84004
Processor optimization removal or modification of security-critical code in AMD Zen allows an authorized attacker to disclose information locally.
-
CVE-2026-102359
Kroki creates diagrams from textual descriptions. Prior to 0.31.2, the TikZ backend's tikz/tikz2svg guard used when KROKI_SAFE_MODE is enabled blocks only \verbatiminput while allowing native TeX file-read primitives including \openin, \read, \input, and \include. TeX Live's permissive openin_any setting remains active, so an unauthenticated actor can submit crafted LaTeX to POST /tikz/svg or the generic diagram API and cause the renderer to read any file accessible to the Kroki container user. The file contents are rendered into the returned SVG, which may disclose mounted secrets, configuration data, credentials, private keys, or process environment data. The -no-shell-escape option prevents shell execution but does not restrict these native file operations. This issue is fixed in version 0.31.2.
-
CVE-2026-120538
The Easy Digital Downloads plugin for WordPress is vulnerable to authentication bypass leading to administrator account takeover in all versions up to, and including, 3.7.2. During a PayPal guest checkout, the email address entered in the guest-checkout field is trusted as proof of identity: edd_get_purchase_form_user() binds the new guest order to an existing WordPress user purely on an email-address match, rewriting the order's user_id to that user's ID. The PayPal capture sink PurchaseUser::maybe_log_in_after_capture() then looked the user up by the order email (get_user_by( 'email', $order->email )) and, when $order->user_id equalled that user's ID (which the email-binding step guarantees), issued an authenticated session with wp_set_auth_cookie(). An unauthenticated attacker who starts a guest checkout using an administrator's email address (where that administrator has no existing EDD customer record) and completes the PayPal purchase flow therefore receives an authenticated session as that administrator — a full account takeover. The same sink is reachable through the standard PayPal checkout and the REST UnbrandedCard and Fastlane flows.
-
CVE-2026-120204
DeepInfer through 0.1.0 contains a path traversal vulnerability in the unauthenticated /profile endpoint that allows remote attackers to create arbitrary directories via the output_dir parameter. Attackers, including cross-site GET requests from an operator's browser, can write chrome trace files to exhaust disk or inodes and use error messages as a filesystem oracle.
-
CVE-2026-118264
Echo is a Go web framework. From 4.0.0 until 4.16.1 and 5.4.1, Echo's Proxy middleware selects a raw tunnel from a client WebSocket upgrade request before verifying that the configured upstream accepted the upgrade. If the upstream rejects the upgrade with an ordinary keep-alive response, later HTTP requests on the connection are sent directly to that upstream without passing through Echo's router or route-specific middleware. This can bypass authentication or authorization enforced only by Echo routes, but requires an accessible proxy route, a cooperating rejected-upgrade response, and a target without independent upstream authorization. This issue is fixed in 4.16.1 and 5.4.1.
-
CVE-2026-118131
GnuPG (version 2.5.21) for Windows contains a vulnerability where default Discretionary Access Control List (DACL) inheritance on the C:\ProgramData directory allows members of the BUILTIN\Users group to create arbitrary subdirectories under this path. An attacker with low-privilege access can create the expected GnuPG configuration directory and inject malicious configuration files consumed by GnuPG components. This can result in code execution within the security context of the victim user.
-
CVE-2026-116818
OpenPanel before 3.0.0 contains an unauthenticated ClickHouse SQL injection vulnerability that allows remote attackers to inject SQL through the projectId parameter of the /live/visitors/:projectId WebSocket, which lacks session and project-access checks. Attackers can use UNION or boolean count oracles to read other tenants' events and profiles, and project members can similarly inject via the cohort.preview timeframe values.
-
CVE-2026-2270
A confused deputy attack exists in the StatefulSet controller that allows a user with namespace-scoped write permissions on StatefulSet and ControllerRevision objects to create a cross-namespace pod. An attacker exploiting this vulnerability would have full control over the resulting pod’s metadata and specification, including namespace selection. Note that the cross-namespace pod will be immediately deleted by the garbage collector unless the attacker is able to construct a valid StatefulSet OwnerReference. This would require referencing the UID of an existing StatefulSet in the victim’s namespace.
-
CVE-2026-105626
HashiCorp Consul and Consul Enterprise are vulnerable to a denial of service in the HTTP API that may allow an unauthenticated attacker to exhaust memory on a Consul agent. This vulnerability (CVE-2026-105626) is fixed in Consul Community Edition 2.1.0 and Consul Enterprise 2.1.0 and 1.21.19.
-
CVE-2026-127960
Denial of service through folded HTTP response headers. The HTTP stream wrapper joins each continuation line of a folded response header to the accumulated value by allocating a new string and copying the whole value, so a response with many short continuation lines takes disproportionate CPU time to parse. An attacker who controls an HTTP or HTTPS response fetched through file_get_contents(), fopen() or other stream functions, by making the application request a chosen URL or by influencing a server it already contacts, can keep PHP workers busy. Repeated or concurrent requests can cause a denial of service. This issue affects PHP versions from 8.1.32 through 8.1.34, from 8.2.28 before 8.2.36, from 8.3.18 before 8.3.37, from 8.4.5 before 8.4.28 and 8.5.* before 8.5.13.
-
CVE-2026-28273
ZoneMinder is a free, open source closed-circuit television software application. Prior to 1.38.4, any authenticated user can use the /api/configs/edit/{name}.json endpoint to modify protected configuration settings because ConfigsController::edit() does not enforce the System=Edit authorization requirement and its system-variable protection is ineffective. An attacker can set ZM_PATH_FFMPEG to a command-bearing value and trigger an exec() sink in web/views/image.php or web/includes/Event.php by requesting frame extraction. This permits arbitrary command execution with the privileges of the web server user and can expose private camera feeds, recordings, and database data or establish persistence. This issue is fixed in version 1.38.4.
-
CVE-2026-21601
This High severity BASM (Broken Authentication & Session Management) vulnerability was introduced in versions 10.0.0 and 11.0.0 of Jira Service Management Data Center. This BASM (Broken Authentication & Session Management) vulnerability, with a CVSS Score of 8.7, allows an unauthenticated attacker to perform actions as another user. Atlassian recommends that Jira Service Management Data Center customers upgrade to latest version, if you are unable to do so, upgrade your instance to one of the specified supported fixed versions: * Jira Service Management Data Center 10.3: Upgrade to a release greater than or equal to 10.3.27 * Jira Service Management Data Center 11.3: Upgrade to a release greater than or equal to 11.3.13 See the release notes (https://confluence.atlassian.com/servicemanagement/jira-service-management-release-notes-780083086.html). You can download the latest version of Jira Service Management Data Center from the download center (https://www.atlassian.com/software/jira/service-management/download-archives). This vulnerability was discovered by Nagli and reported via our Bug Bounty program
-
CVE-2026-124666
Authentication bypass by spoofing vulnerability in Apache APISIX. On a route where multi-auth falls back from wolf-rbac to another successful authenticator, client-supplied Wolf identity headers can remain in the request. A downstream service that relies on those documented headers may therefore receive an identity that was not established by the successful authenticator. This issue affects Apache APISIX: from 3.8.0 through 3.19.0. Users are recommended to upgrade to version 3.20.0, which fixes the issue.
-
CVE-2026-17343
IBM API Connect 12.1.0.0 through 12.1.1.2 could allow a remote authenticated attacker to bypass authorization checks to access or modify sensitive information due to improper authorization.
-
CVE-2026-63482
Yeti bridges the gap between CTI and DFIR practitioners by providing a Forensics Intelligence platform and pipeline. Prior to 2.5.1, YETI registers the audit router without Depends(auth.get_current_active_user), and core/web/apiv2/audit.py does not perform a route-level authentication check for /api/v2/audit/timeline/{id}. An unauthenticated network attacker can use an empty, partial, or specific wildcard path to query TimelineLog records containing the actor, target_id, action, and details fields, including analyst usernames, object identifiers and values, and before-and-after object data. This exposes the operational history of threat-intelligence objects while other application routers require authentication. This issue is fixed in version 2.5.1.
-
CVE-2026-114582
An issue was discovered in Django 6.1 before 6.1.3, 6.0 before 6.0.10, and 5.2 before 5.2.19. `django.contrib.gis.gdal.SpatialReference` (and related `transform()` methods and `srs`/`srid` setters) do not restrict GDAL’s file and network access when parsing string input, which allows attackers to trigger server-side requests or local file reads via crafted spatial reference strings. 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 Adraca for reporting this issue.
-
CVE-2026-78192
WebDyne versions from 2.022 before 3.022 for Perl disclose application source to unauthenticated clients via the viewer in the default index page. WebDyne's wrapper scripts, such as webdyne.psgi, enable the built-in index page by default. Its source viewer returns files under the document root named in the source parameter of an HTMX request, including .psp, .pm and .pl files. A remote attacker can read the application's Perl code, including the __PERL__ blocks of .psp pages and any credentials they hold.
-
CVE-2026-119191
Forge (also called node-forge) is a native implementation of Transport Layer Security in JavaScript. Prior to 1.4.1, the affected implementation used asn1.fromDer(), BIT STRING decoding, maxDepth to process the relevant input or security decision. node-forge decoded ASN.1 stored inside BIT STRING values without consistently applying the configured maxDepth limit and recursively retained copies at each nesting level. Crafted DER containing nested BIT STRINGs could cause excessive recursion, CPU use, and quadratic memory consumption during asn1.fromDer(). This issue is fixed in version 1.4.1.
-
CVE-2026-103093
ServiceNow has remediated a sandbox escape security issue that was identified in the ServiceNow AI Platform. This vulnerability could enable an authenticated user with the admin role to bypass Rhino sandbox restrictions and execute arbitrary code on the underlying host operating system, potentially enabling further unintended access. ServiceNow deployed an update to hosted instances, and ServiceNow provided the update to our partners and self-hosted customers. We are not currently aware of malicious exploitation against ServiceNow instances. We recommend customers promptly apply appropriate updates or upgrade to a patched release if they have not already done so.
-
CVE-2026-119050
Agno through 3.1.11 contains an unsafe reflection vulnerability in PandasTools that allows attackers to invoke arbitrary pandas functions by supplying model-controlled function names to getattr(). Attackers can use prompt injection to make the agent call pandas.read_pickle on attacker-controlled paths for code execution, or call other readers to read local files.