
Magento Zero-Day, Record Patch Tuesday and AI Assistant Exfiltration
Coverage: 08 Sept 2026 (UTC)
< view all daily briefs >The September 8, 2026 batch combines an actively exploited Magento zero-day, a record-scale Microsoft Patch Tuesday and a set of cloud platform changes focused on post-quantum key exchange, mutual authentication and compliant data retention. Research into AI assistants shows covert channels, reusable encrypted reasoning traces and agent-driven credential theft moving into demonstrated attacks. Mobile and GPU hardware research, together with a new French government incident response mechanism, completes the day.
Cloud Platforms Harden Encryption, Authentication and Retention
Cloudflare announced Automatic Key Exchange, an enhancement to Automatic SSL/TLS that probes origin servers to determine which TLS 1.3 key agreement algorithm to lead with. Previously, Cloudflare always sent an X25519 keyshare first, which caused HelloRetryRequests and added latency when origins preferred other algorithms, notably the post-quantum hybrid X25519MLKEM768. According to Cloudflare, the rollout cut HelloRetryRequests from about 52% to 3.7% and reduced median handshake latency by roughly 150 ms at the p90, while enabling hundreds of thousands of domains to use post-quantum origin connections without manual configuration. Of more than a million domains assigned preferences so far, about 64% remain on X25519 and about 33% now prefer X25519MLKEM768, with roughly 9,000 domains updated each day. The feature is enabled by default for TLS 1.3 origins. A Compliance requirements setting can restrict which algorithms are advertised or negotiated, although Cloudflare cautions that overly strict enforcement, such as requiring post-quantum hybrids on origins without compatible support, can break connections. The company presents the change as a way to reduce harvest-now, decrypt-later risk with minimal operator intervention.
On AWS, API Gateway mTLS support now extends to REST API backend integrations. API Gateway can present an AWS Certificate Manager (ACM) certificate to backend endpoints during the TLS handshake, replacing the earlier generated self-signed certificate, so customers can use certificates from their own trusted PKI or from AWS Private Certificate Authority. Certificate imports, renewals and reimports propagate automatically without redeployments or downtime, and the capability is available in all commercial and AWS GovCloud (US) Regions where REST APIs are supported. Separately, S3 Object Lock gains variable retention through event holds, which allow WORM protection to begin from a future triggering event such as a contract closing or the completion of an audit. Unlike legal holds, which end protection as soon as they are lifted, an event hold enforces its defined retention period after release. New IAM and bucket policy condition keys control who can set or release holds and enforce minimum and maximum durations, hold actions are recorded in AWS CloudTrail, and the feature is available in all AWS Regions, including China Regions, at no additional charge. AWS notes that Cohasset Associates assessed the capability for environments subject to SEC Rule 17a-4(f), FINRA Rule 4511 and CFTC Regulation 1.31.
Two further AWS changes address lifecycle risk and data handling. Amazon EMR will continue to provide critical security fixes on a best-effort basis for releases 5.36 and 6.6 through 6.15 until June 30, 2027, and will extend Standard Support for releases 7.0 through 7.10 until August 31, 2027, for customers actively migrating. Coverage applies to EMR on EC2, EMR on EKS and EMR Serverless in all Regions where EMR is offered; customers must contact AWS Support with a migration plan, and there is no opt-in fee. Bedrock AgentCore Memory adds an IngestData API that pushes conversational or generic JSON content, such as behavioral events, activity logs and system events, directly into long-term memory extraction without first creating a short-term event. Extraction status and results can be inspected with ListMemoryRecords and RetrieveMemoryRecords, real-time notifications can be streamed via Kinesis, and failed extractions can be retried or diagnosed with ListMemoryExtractionJobs.
Exploited Zero-Days Set Patch Priorities
Adobe published emergency updates for a Magento zero-day, CVE-2026-75650, a critical remote code execution flaw rated CVSS 10.0 that affects Adobe Commerce, Adobe Commerce B2B and Magento Open Source. Tracked as "StyleSmuggler" by Sansec, the bug was first observed exploited in the wild on September 4, 2026, and abuses Magento template processing to inject PHP when generating a "Payment Transaction Failed Reminder" email. Exploitation has been used to deploy a Rust-based Linux backdoor and PHP droppers that install web shells. Adobe released the VULN-39341 hotfixes and instructs administrators to apply the patch for their version and rotate encryption keys. On September 8, CISA added the flaw to its Known Exploited Vulnerabilities catalog, mandating fixes for federal agencies by September 11, 2026. Adobe also patched more than 170 additional vulnerabilities, including high-severity issues in Campaign Classic and ColdFusion, with no observed exploitation of those flaws to date.
Microsoft's September Patch Tuesday, as analyzed by Cisco Talos, addresses 973 vulnerabilities, 113 of which are rated critical, including 82 remote code execution flaws. Two elevation-of-privilege bugs have been exploited in the wild: CVE-2026-81963 in the Windows Update Stack and CVE-2026-85880 in ALPC, both scored CVSS 7.8. Microsoft flagged several issues as more likely to be exploited, including high-severity flaws in Kerberos, RRAS, Windows DNS Server and Azure Cosmos DB with CVSS scores up to 9.8. Many of the flaws stem from heap- and stack-based buffer overflows, use-after-free conditions, integer overflows and improper input validation. Talos advises prioritizing Internet-facing services, domain controllers and widely used Office and server products, and monitoring for exploitation indicators tied to the high-profile CVEs.
Red Hat disclosed a critical FreeIPA flaw chain that allows a client that has never authenticated to create a Kerberos identity of its choosing and end up in the administrators group on an otherwise default FreeIPA deployment. The chain combines a FreeIPA access-control rule that lets users manage one-time-password tokens without prior login (CVE-2026-76578, preliminary CVSS 9.8) with a 389 Directory Server bug that treats an unauthenticated client as having an empty name that matches ownership checks (CVE-2026-76560, CVSS 7.5). FreeIPA 4.13.4 contains the FreeIPA-side fix, and 389-ds patches are being distributed through vendor advisories; Red Hat reports no exploitation in the wild. A separate important flaw, CVE-2026-79678 in the idp-add command, can disclose environment variables and cause memory exhaustion, a particular concern for containers that may retain initial passwords in the process environment. Interim mitigations include restricting LDAP access on ports 389 and 636 and disabling anonymous binds. The advisories do not state whether fixes remove previously injected identities, and no detection indicators have been provided.
AI Assistants and Agents Emerge as Attack Surface
Check Point Research (CPR) found a covert two-way channel between ChatGPT code execution containers belonging to separate accounts, detailed in its ChatGPT Gmail research. The channel abused an internal JFrog Artifactory instance used for fetching packages, whose item properties any container could write and read, effectively forming a shared clipboard between supposedly isolated sessions. In a proof of concept, a planted instruction delivered through a pasted prompt, a shared conversation link or a custom GPT caused a victim's session to pull data from the connected Gmail account and deliver it to an attacker's account, all during an ordinary-looking reply. The only visible sign was a small "Talked to Gmail" label. OpenAI confirmed the Artifactory instance has been decommissioned and no user action was required, though CPR notes that the underlying architectural pattern remains a general risk, describing affected assistants as potential coerced insiders. CPR recommends inventorying AI tools and their connections, deploying runtime protections against prompt injection and data exfiltration, and enforcing governance over what AI agents are allowed to do.
Google Threat Intelligence Group's AI Threat Tracker describes adversaries moving from basic prompting to agentic workflows and automated toolchains that compress defenders' response windows. In Q2 2026, GTIG observed one agent-enabled campaign complete compromise, planning, build and mass credential harvesting in under six hours. Financially motivated actor UNC6780 (TeamPCP) conducted large-scale open source supply chain compromises across PyPI, npm and Docker Hub, using the DUSTMAKER credential stealer to harvest developer tokens, publish trojanized packages and bypass AI agent trust checks. DUSTMAKER includes features aimed at AI tooling, including hidden project directories, configuration hijacking for persistence, prompt injection to coerce AI assistants, CI/CD masquerading with log deletion, and payloads designed to blunt LLM security scanners. GTIG also reports that proprietary models, prompts, source code, API credentials and cloud compute quotas are being exfiltrated or co-opted. Google says it has disrupted malicious infrastructure, hardened models and classifiers, updated mitigations for open source supply chain risks and operationalized autonomous defenses.
Research highlighted by Bruce Schneier on stealing reasoning traces found that encrypted chain-of-thought blocks returned to clients by major LLM providers are interchangeable across sessions, users and even different models within the same provider ecosystem. Injecting a target model's encrypted trace into a weaker, less-restricted model allowed the latter to output the hidden reasoning in plaintext. The researchers decoded more than 315,000 blocks from publicly shared session logs, recovering hundreds of PII artifacts and credentials, and also demonstrated bypass of anti-distillation protections, exposure of otherwise suppressed hazardous content and silent prompt injection through payloads embedded in encrypted blocks. The technique was validated across several vendors and responsibly disclosed. Recommended fixes include binding encrypted traces to specific sessions and attested runtime contexts, encryption schemes that prevent cross-model reuse and clearer developer guidance on sharing session artifacts.
On the deployment side, AWS made OpenAI's GPT-6 Astra generally available on Amazon Bedrock. The model supports up to 1 million input tokens and advanced computer and browser use, with highlighted use cases including autonomous agents, large-scale document analysis and complex software investigation. AWS points to established controls for securing workloads, governing access and auditing model invocations, while OpenAI is releasing enterprise plugins for ChatGPT Work that extend Astra's browser capabilities across common business applications.
Mobile Worms, GPU Rowhammer and French State Response
Researchers at security firm Calif demonstrated a WeChat zero-click worm that takes over a WeChat account when the target receives a call from an existing contact. In the demonstration, an Android phone called an iPhone and took control of its WeChat account while it was still ringing, after which the compromised iPhone called a second Android device to propagate the compromise. The worm grants full control of the WeChat account, including messages, calls and account-linked services, but does not on its own take control of the underlying phone. Calif reported the issue to Tencent in July; according to Calif, Tencent released app updates in late August and applied a server-side block that mitigated the exploit for all users by August 28. Tencent has not published an advisory or CVE, and Calif disclosed that it used AI tooling to discover the vulnerability and accelerate exploit development. Users are still advised to run current app versions.
Dark Atlas detailed the THost9 Android RAT, a packed Android RAT cluster grouped under the artifact label Hagaseca. The loader hides code in an embedded asset decoded with single-byte XOR and gzip before loading the second-stage payload tc9.dex, which provides shell execution, file transfers, tunneling, a reverse shell and modular components. An ADB-targeting worm discovers ADB services, authenticates with prepared ADB keys and installs the malware on reachable devices, and newer samples include a Frida anti-analysis check. Public incident reports tie THost9 and an earlier THost4 variant to infections of Android phones and Redroid containers with exposed ADB from October 2024 through 2026. Recommended mitigations include removing public ADB exposure, reviewing accessibility service grants and cleaning persistent Redroid data; the cluster is not attributed to a verified threat actor.
University of Toronto researchers introduced GPUThor, a Rowhammer-class attack on Nvidia Ampere GPUs with GDDR6 memory. After finding that Target Row Refresh triggers only once every 72 refresh cycles, the team used an uneven, aggressive hammering pattern to achieve between 72,000 and 377,000 bit flips per gigabyte, enough to produce double- and triple-bit errors that overwhelm single-bit ECC correction. Tests on A4000, A4500, A5000 and A6000 cards caused accelerator reboots and device failures consistent with denial-of-service impact; arbitrary code execution was not demonstrated. The findings bear on shared GPU resources in cloud and AI environments.
In France, the national cybersecurity agency ANSSI announced on September 7 the creation of REACTIV, an interministerial incident response and action mechanism for state services. It allows ANSSI to compel ministries, within tight deadlines, to implement immediate measures protecting citizens' data held by government administrations, and to centralize and lead technical crisis communications during attacks on state services. The move follows an attack on the Directorate-General for Public Finance (DGFiP) that exposed sensitive data of between 350,000 and 678,000 taxpayers and led to the arrest of two youths suspected of participating as members of the "ZeroBytes" group. ANSSI has not disclosed the human or financial resources assigned to REACTIV.