Over the past two weeks, the global technology sector has been hit by a string of major security incidents. They include Binance losing more than $6.3 billion to North Korean hackers, a maliciously altered Claude Code extension for Cursor, and newly disclosed high-risk flaws in core software such as PHP, GitLab and Linux. As AI agents gain the ability to run code and access enterprise systems on their own, existing security problems are getting amplified.
Jack Cable, security product lead at AI coding company Cognition and co-founder of Corridor, published a post on Nov. 4 titled A Practical Guide to Security in November 2026. In it, he reviewed the recent incidents and set out security steps companies should take immediately and within the next 30 days. Cable wrote that AI does let attackers find and exploit vulnerabilities faster and at lower cost, but basic security controls can still shut down many attack paths.
Binance breach traced back to malware on an engineer’s machine
Cable opened with the recent Binance hack, describing it as an incident that has been characterized as the largest theft on record. The initial entry point, he said, was not an advanced cryptographic attack. It was a personal work device used by an engineer.
Citing Binance’s account, the post said several engineers were targeted through fake job interviews and tampered software installers, which infected their laptops with malware. A deployment token stored on one of those devices then gave the attackers a path to Binance’s wallet signing service.
Cable argued that even if an engineer’s computer is compromised, that should not give an attacker access to the production environment, and it should not end in multi-billion-dollar asset losses.
For him, the core failure was weak separation between employee devices, development environments and critical production systems. Once high-privilege credentials are left on a standard developer machine, a relatively cheap social-engineering campaign can open access to infrastructure that should sit behind several layers of protection.
Claude Code extension backdoored, with auto-update turning into an attack path
Cable also pointed to software supply-chain risk in AI coding tools. He said Cursor disclosed on Oct. 29 that an attacker had published a maliciously modified Claude Code extension for Cursor. The goal was to install a backdoor on developers’ machines that could maintain persistence and allow remote access over SSH.
That case stood out for another reason. Editor extensions often update automatically, so developers may receive a malicious version during a normal update cycle even if they never knowingly install suspicious software.
According to Cursor, fewer than 5,000 users installed the affected extension. The tainted version was not the same as the one published on the Open VSX extension registry, which suggests the compromise may have happened later in the release or distribution chain.
The incident also highlights a newer risk around AI coding stacks. Developers now use not only traditional packages, but also editor extensions, MCP servers and AI agents with terminal and filesystem access. If any one of those layers is compromised, an attacker may inherit the system permissions the developer already has.
Cursor said the affected versions were 2.1.312 and 2.1.313, mainly on Windows and macOS. The malware attempted to establish an SSH backdoor, but some steps may not have executed successfully because the code itself contained errors and most users did not launch Cursor with administrator privileges.
Cursor has removed the malicious versions, shut down the server used by the attacker, and advised affected users to remove the backdoor, inspect persistent processes and rotate related credentials.
PHP, GitLab and Linux flaws add to the pressure, while a certificate authority used “4” as a private key
Beyond those two incidents, Cable listed several other serious vulnerabilities disclosed recently, showing how broadly the risk now stretches across web infrastructure, code management platforms, virtualization environments and operating systems.
One of the most closely watched cases was PHP2Shell. Cable said the flaw stems from a memory-safety issue in the PHP runtime. Under certain configurations, an attacker may be able to trigger it with a single network request and without logging in, potentially affecting millions of websites.
GitLab was also reported to have a remote code execution flaw that can be triggered without authentication, raising the prospect of compromise at systems that hold critical enterprise source code. Edge devices directly exposed to the internet, including VPNs and firewalls, remain prime targets for state-backed hacking groups.
In another case, Indian digital certificate authority eMudhra was said to have set its private key to the number “4,” effectively stripping the key of secrecy and creating the possibility that a third party could impersonate it and issue certificates in its name.
Cable also noted that Vercel’s Sandbox bug bounty program found a guest-to-host escape flaw in KVM virtualization, creating a path for code running inside a virtual machine to gain root access on the host. He added that CVE-2026-98432 in the Linux TCP stack could allow remote privilege escalation under certain Multipath TCP configurations.
What companies should do now and within 30 days
Cable said the answer is not to rip out existing systems wholesale. The priority is to build controls that reduce the attack surface. He split the recommendations into two phases: actions to take immediately and actions to complete within 30 days, covering enterprise accounts, development environments, source code, third-party dependencies and infrastructure monitoring.
On account security, he recommended phishing-resistant multi-factor authentication for Google Workspace, GitHub, cloud services and identity platforms, using passkeys or hardware security keys instead of relying only on SMS. Privileged accounts should also be limited to logins from company-managed devices.
He also said companies should roll out password managers and ad blockers across the organization, and review all internet-facing domains and services. Cable specifically linked this to the Binance attack path, which began when an engineer searched for an AI coding tool and downloaded a tampered installer. Malicious search ads and fake download sites, in his view, belong inside the threat model.
Within the next 30 days, companies should deploy mobile device management, or MDM, and endpoint detection and response, or EDR, on all work devices. They should also create allow lists for browser and code editor extensions and consider delaying automatic updates to reduce exposure to supply-chain attacks.
AI agents are changing development, and sensitive credentials should not stay on developer laptops
Cable put special emphasis on how AI coding tools are changing the structure of developer risk. In the past, engineers mainly worked through editors, terminals and package managers. Now they may also run multiple MCP servers, AI agents and third-party extensions at the same time.
Those tools can do more than read code. Some can operate terminals, modify files and execute system commands on their own. If a package, extension or AI agent falls under an attacker’s control, it may inherit all the access the developer already has.
He recommended that companies immediately use tools such as TruffleHog to scan developer devices for cloud credentials, API keys, SSH keys and secret .env files, and to inventory installed extensions, plugins and MCP servers.
Over a longer horizon, he said organizations should centralize secrets in a secrets manager and gradually move toward cloud coding agents and remote development environments. Those setups can assign the minimum necessary permissions for a task and are easier to monitor, isolate and rebuild. Compared with running AI agents directly on personal developer devices, centrally managed remote environments can reduce the damage from a single compromised machine, though Cable said permissions and network isolation still need to be designed carefully.
Use AI for code scanning, but feed it the company’s threat model
Cable did not argue that companies should scale back AI in software development because AI has increased security risk. He took the opposite view. He said businesses should use frontier models to scan internal codebases and important open-source dependencies, with priority on internet-exposed services and systems tied to authentication, payments and customer data.
He also warned that asking a general-purpose model to “find bugs” can produce a large number of alerts with little practical value. Companies should provide their own threat model to the scanning system so the AI understands how the code is deployed, which services are exposed to the internet, and which data and systems matter most.
For third-party software supply chains, Cable recommended delaying installation of new releases from package managers such as npm and PyPI by at least seven days, reducing the chance of pulling a malicious version before it is identified.
He also said companies should pin dependency versions, commit lockfiles and require continuous integration systems to install packages from those locked versions rather than fetching the latest release on every build.
On infrastructure monitoring, Cable recommended complete logging for identity systems, cloud operations, code management and access to production data, with records protected so an attacker cannot easily delete them even after gaining production access. Alerts should fire when there is a new administrator, a change to MFA settings, a large data export or unusual access to sensitive credentials.


