Daily Cyber Briefing — November 13, 2026
The bugs AI finds are now in real attack chains, and defenders are being handed the same tools while businesses decide whether to let agents in at all. Window roughly Nov 12, 7:00 a.m. to Nov 13, 7:00 a.m. CT. Sources are linked. Verification is marked on each item.
PHP2Shell chain breaks into Check Point gateways, plants a backdoor, and wipes ESXi hosts
Verified
Check Point Research says an unknown actor began exploiting PHP2Shell (CVE-2026-109618 and CVE-2026-109619) on unpatched Mobile Access portals of Check Point Security Gateways on October 30, about a week after a fix was available, first against Israeli organizations and then in a broader, opportunistic wave on November 2. The attackers escalated to root through a previously unreported Gaia OS flaw in the DynamicCLI interface, CVE-2026-123039 (CVSS 7.8), which Check Point says it saw exploited against "a couple of dozen customers." They installed PHP web shells and a gateway implant called TSGA, with a TSProxy component that hides attacker connections inside normal appliance traffic and can tamper with firewall logs. Some first-wave victims then saw the attackers move over SSH to VMware ESXi hosts, run a wiper called destroy_v6, and leave politically themed notes.
PHP2Shell is the PHP flaw that a researcher at Calif found with an OpenAI model in October. This is the first detailed chain showing it used to take over security appliances and destroy data.
Why it matters: A firewall that gets popped becomes the attacker's quietest foothold. If you run Check Point gateways with Mobile Access on R81 through R82.20, patch both PHP2Shell and CVE-2026-123039, hunt for the indicators Check Point published, and treat a compromised gateway's configs and logs as untrustworthy.
Sources: Check Point Research · Check Point advisory · The Hacker News
OpenAI ships Daybreak Red Team and Defense Factory, and previews a stronger cyber model
Verified
At its Cyber at the Frontier event on November 12, OpenAI put Daybreak Red Team into early access. It runs automated penetration tests and exploit validation using OpenAI's frontier cyber models. OpenAI also announced Daybreak Defense Factory, which traces a vulnerability to the affected code, finds the right code owner, and prepares a patch for review. OpenAI previewed an upcoming cyber model it says scores 77.0% on ExploitGym and 79.1% on PatchBench; those figures are OpenAI's own. Greg Brockman said coordinating the PHP2Shell disclosure took about two months and consumed the bandwidth of OpenAI, partner companies and governments, and that the approach won't scale to what models find next.
Why it matters: Continuous AI pentesting is becoming a product any team can buy, which raises the bar for what counts as "tested." The harder problem is downstream: patching and disclosure pipelines built for a handful of critical bugs a month now have to absorb far more.
Sources: OpenAI · Livestream · SecurityWeek
N-central exploits hit 21 servers, 19 of them already on Hotfix 2
Verified
Exploitation of N-able N-central's 4EverAble2Shell chain has moved past probing. Previdian tracks the unauthenticated chain, which combines an Apache Axis flaw (CVE-2026-90436) with an N-central privilege escalation, and saw attackers fetch a shell script from 157.20.241[.]16. Huntback reports the same address installed XMRig miners on 21 N-central servers, 19 of which were running Hotfix 2, according to The Hacker News. N-able says it is restricting some N-central features and integrations on hosted environments while it remediates, and tells on-premises customers to upgrade Agents and Probes to 2026.4 HF1 or later and disable the Legacy API.
Why it matters: A coin miner is the mild outcome on a server that can push commands to every managed endpoint. Keep on-premises N-central off the internet until a full fix ships, and check for jup.sh and unexpected CPU load.
Sources: The Hacker News · Previdian
United will block AI agents from managing reservations
Verified
United Airlines said it will block third-party AI agents from managing customers' reservations and purchases, on its own platforms and through aggregators and agencies. It cited the risk of leaking credentials users hand to agents, unintended purchases or use of MileagePlus miles, and FAA and Department of Transportation rules that require a human to acknowledge what's in checked bags. The Record notes Bank of America recently told customers never to share their logins with agents.
Why it matters: Large consumer brands are answering agentic AI by shutting the door, because they have no way to tell a delegated agent from an attacker holding the same password. Expect pressure for scoped, verifiable delegation instead of shared credentials.
Sources: The Record
SignDrop: public Windows CLFS exploit reaches SYSTEM with no patch
Verified
A researcher publishing as MSNightmare released proof-of-concept code on November 10 that abuses the log authentication Microsoft added to the Common Log File System driver in 2025. When a privileged process opens a prepared log (Defender's update interface in the published code), the driver creates a companion file as SYSTEM in a location the user controls, which can be chained to a SYSTEM shell. It was tested on Windows 11 26H2 and Server 2025 builds, and default enforced configurations are in scope. Microsoft has not confirmed it or assigned a CVE. 78researchlab says it works but not reliably enough for operational use.
Why it matters: Public local privilege escalation code for current Windows tends to show up in ransomware kits quickly. Watch for unusual CLFS file creation by Defender and keep EDR tamper protection on; don't disable CLFS authentication as a workaround.
Sources: The Hacker News
Also on the wire
- Operation Endgame took down Gootloader, CastleRAT and StealC infrastructure, seizing 235 servers and 257 domains and securing about 6.4 million stolen logins. Europol
- The Justice Department moved to seize or restrain more than $1.2 billion of the Binance theft it attributes to North Korea's TraderTraitor. The Record
- A critical Red Hat Quay flaw, CVE-2026-123710, lets an attacker plant a trusted signing key and read or overwrite every private image, anonymously in default setups. Red Hat
- Island found 186 lookalike domains for 27 AI brands, including ChatGPT, Claude, Gemini and Cursor, pushing ClickFix lures and search-hijacking extensions. The Hacker News
- Socket says Japan logged a record 233 disclosed cyber incidents in October, led by web-system breaches. Socket on X
Cyber covers AI and agentic threats, supply-chain and developer-tool compromises, and how enterprise security absorbs them. Signal stays the AI research stream. ← All bulletins