Security Digest — 2026-05-16

Today’s headlines split between an active on-prem Exchange zero-day under exploitation and a flood of new research into the security of LLM agents, their third-party skills, and their supply chains. AI is showing up on both sides of the fence: Microsoft is using fleets of agents to hunt Windows bugs, while researchers keep finding new ways those same agents can be tricked, fingerprinted, or co-opted.


AI Security Research

The Great Pretender: A Stochasticity Problem in LLM JailbreakArXiv cs.CR. The authors argue that the dominant jailbreak-attack benchmarks systematically overstate success rates because they ignore the inherent stochasticity of LLM sampling, calling much of the published jailbreak literature into question.

WARD: Adversarially Robust Defense of Web Agents Against Prompt InjectionsArXiv cs.CR. Existing guard models for web agents collapse under adaptive prompt-injection attacks embedded in HTML or visual interfaces; WARD proposes an adversarially trained defense aimed at closing that gap.

AgentTrap: Measuring Runtime Trust Failures in Third-Party Agent SkillsArXiv cs.CR. As third-party “skills” become the de facto package ecosystem for LLM agents, this work introduces a framework to measure how often skills violate trust boundaries at runtime — a precursor to treating skill marketplaces as a real attack surface.

Exploiting LLM Agent Supply Chains via Payload-less SkillsArXiv cs.CR. A companion line of work shows that malicious agent skills don’t need executable payloads at all: natural-language instructions and templates alone are enough to hijack agent behavior across an open marketplace.

ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity AgentsArXiv cs.CR. Rather than treating “did it crash?” as exploitation success, ExploitBench grades agents along a ladder from triggering a buggy line to achieving full target control, giving a more honest picture of offensive LLM capability.

RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL AttacksArXiv cs.CR. Prior watermarking schemes claimed robustness to paraphrasing; RLCracker trains an RL attacker that adaptively erases watermarks, exposing those robustness claims as fragile in the adversarial worst case.

“Tab, Tab, Bug”: Security Pitfalls of Next Edit Suggestions in AI-Integrated IDEsArXiv cs.CR. Next Edit Suggestion features in IDEs construct rich context from across the workspace, and the authors show this enables novel injection paths where a single tainted file can steer the assistant into emitting vulnerable edits elsewhere in the project.

Capacitive Touchscreens at Risk: A Practical Side-Channel Attack on Smartphones via Electromagnetic EmanationsArXiv cs.CR. TESLA, a contactless EM side-channel attack, is shown to recover touchscreen input from smartphones without invasive probes, raising the bar for what counts as a “physical-only” threat model.

Bypassing On-Camera Age-Verification ChecksSchneier on Security. AI-based video age-verification systems can be defeated with a fake mustache — a small but pointed reminder that biometric gating is only as strong as its weakest visual feature.

Microsoft pits more than 100 AI agents against each other to find Windows vulnerabilitiesThe Decoder. Microsoft’s MDASH system used over 100 specialized agents to surface 16 Windows security flaws on a single Patch Tuesday, four of them critical — though it declined to disclose which models power the swarm.

The Boring Stuff is Dangerous NowDark Reading. AI agents that can hunt and exploit obscure vulnerabilities are arriving at the same time as a tide of AI-generated code that may itself be flawed, forcing defenders to rethink what “low-priority” bugs even mean.

Vulnerabilities & Exploits

On-Prem Microsoft Exchange Server CVE-2026-42897 Exploited via Crafted EmailThe Hacker News. Microsoft has disclosed an actively exploited spoofing/XSS vulnerability (CVSS 8.1) in on-premise Exchange that lets attackers run arbitrary code against Outlook-on-the-web users via a crafted email. Mitigation guidance is also out from Microsoft for the same flaw.

Microsoft Exchange, Windows 11 hacked on second day of Pwn2OwnBleepingComputer. Day two of Pwn2Own Berlin 2026 produced 15 unique zero-days against Windows 11, Microsoft Exchange, and Red Hat Enterprise Linux for Workstations, with $385,750 in awards paid out.

Popular node-ipc npm package compromised to steal credentialsBleepingComputer. Malicious versions of the widely used node-ipc package were published to npm with credential-stealing payloads, marking yet another high-impact supply-chain compromise in the JavaScript ecosystem.

TanStack Supply Chain Attack Hits Two OpenAI Employee Devices, Forces macOS UpdatesThe Hacker News. OpenAI disclosed that the “Mini Shai-Hulud” supply chain attack on TanStack reached two corporate devices; the company says no user data, production systems, or IP were affected, but the incident triggered company-wide macOS updates.

NGINX Remote Code Execution Vulnerability DisclosedGigazine. Researchers at DepthFirst found four memory-corruption issues in NGINX, including CVE-2026-42945 — a heap buffer overflow that traces back to 2008 and went unpatched for nearly 18 years before being caught.

Funnel Builder WordPress plugin bug exploited to steal credit cardsBleepingComputer. A critical flaw in the Funnel Builder plugin is being abused in the wild to inject JavaScript skimmers into WooCommerce checkout pages. A separate set of Avada Builder plugin flaws — installed on ~1M sites — also enable arbitrary file reads and database extraction.

Turla Turns Kazuar Backdoor Into Modular P2P Botnet for Persistent AccessThe Hacker News. The Russian state-sponsored Turla group has re-engineered its long-running Kazuar backdoor into a modular peer-to-peer botnet, trading central infrastructure for stealth and persistence on compromised hosts.

Four OpenClaw Flaws Enable Data Theft, Privilege Escalation, and PersistenceThe Hacker News. Cyera disclosed “Claw Chain,” a set of four flaws in OpenClaw that can be combined to steal data, escalate privileges, and persist — notable because OpenClaw is increasingly being adopted as an agentic backbone.

Taiwan Bullet Train Hack Highlights Cybersecurity Gaps in Rail SystemsDark Reading. A student using software-defined radio shut down three Taiwanese bullet trains for nearly an hour, triggering an anti-terrorism response and exposing how fragile rail control systems remain to entry-level RF attacks.

Policy & Compliance

CISA Adds Cisco SD-WAN CVE-2026-20182 to KEV After Admin Access ExploitsThe Hacker News. CISA added a newly disclosed Cisco Catalyst SD-WAN Controller flaw to its Known Exploited Vulnerabilities catalog, putting federal agencies on a mandatory remediation clock.

Congress Puts Heat on Instructure After Canvas OutageDark Reading. The House Homeland Security Committee sent a formal letter to Instructure on the same day the company said it had reached “an agreement” with the ShinyHunters extortion crew behind the Canvas cyberattack — escalating scrutiny of edtech incident handling.

Microsoft backpedals: Edge to stop loading passwords into memoryBleepingComputer. After previously calling the behavior “by design,” Microsoft will change Edge so it no longer loads saved passwords into process memory in cleartext at startup — a reversal driven by sustained public pressure.

Microsoft to automatically roll back faulty Windows driversBleepingComputer. Microsoft is rolling out a capability to remotely revert problematic drivers delivered through Windows Update, a clear post-CrowdStrike posture shift toward platform-level rollback as a baseline safety control.