# How a Poisoned Security Scanner Became the Key to Backdooring LiteLLM | Snyk | Garett Moreau 🇺🇸 **By:** Garett Moreau 🇺🇸 **Published:** 2026-03-25T10:25:46.768Z **Source:** [LinkedIn](https://www.linkedin.com/posts/garettm_weaponized-malicious-code-someone-just-share-7442517077254586368-ixX6) --- # Garett Moreau 🇺🇸’s Post [Garett Moreau 🇺🇸](https://www.linkedin.com/in/garettm?trk=public_post_feed-actor-name) WEAPONIZED MALICIOUS CODE: "Someone just poisoned the Python package that manages AI API keys for NASA, Netflix, Stripe, and NVIDIA. 97 million downloads a month...and a simple pip install was enough to steal everything on your machine. The attacker picked the one package whose entire job is holding every AI credential in the organization in one place. OpenAI keys, Anthropic keys, Google keys, Amazon keys - all routed through one proxy. All compromised at once." Apparently the poisoned version was published straight to PyPI.. no code on GitHub.. no release tag.. no review. Just a file that Python runs automatically on startup. You didn’t need to import it. You didn’t need to call it. The malware fired the second the package existed on your machine. The attacker vibe coded it. The malware was so sloppy it crashed computers. And used so much RAM a developer noticed their machine dying and investigated. They found LiteLLM had been pulled in through a Cursor MCP plugin they didn’t even know they had. If the code had been cleaner nobody notices for weeks. Maybe months. The payload was 3 stages, deploy privileged containers across every node in the cluster and installed a persistent backdoor waiting for new instructions. TeamPCP posted on Telegram after: “Many of your favorite security tools and open-source projects will be targeted in the months to come, stay tuned.” Every AI agent, copilot, and internal tool your company shipped this year runs on hundreds of packages exactly like this one… [https://lnkd.in/gxJP9g4q](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgxJP9g4q&urlhash=bm2-&trk=public_post-text) [ How a Poisoned Security Scanner Became the Key to Backdooring LiteLLM | Snyk snyk.io ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fsnyk%2Eio%2Farticles%2Fpoisoned-security-scanner-backdooring-litellm%2F&urlhash=MzIr&trk=public_post_feed-article-content) This is exactly why “we use open source” is not a security strategy. The minimum bar now is locked dependencies, SBOM, and an internal package proxy so random upstream changes don’t hit prod. AI agents just accelerate the blast radius because they pull in libraries faster than humans review. If a team can only do one thing this quarter, what’s the first control you’d mandate, signed builds, dependency allow lists, or proxying everything? The part that should keep every security team up at night is how the discovery happened: a developer noticed their machine dying, not any detection tool. If the attacker had written cleaner code, this runs silently for weeks across organizations that trust their AI toolchain implicitly. Supply chain attacks targeting the AI dependency stack are going to accelerate because the attack surface is enormous and most organizations treat pip install the same way they treat flipping a light switch. The Cursor MCP plugin angle is especially concerning since developers may not even realize what dependencies they are pulling in. When you skip the basics, it always comes back to bite you. Don’t avoid the new thing, because you may miss key opportunities…. BUT, make sure you limit the blast radius if something goes bad. Segment it off while you learn about how it works, limit the amount of access it has (read only, small scope at first), limit outbound access, human reviews before key changes are made (build trust in automation), make sure it is logging actions and that is reviewed. Otherwise, these type of supply chain attacks are going to steal valuable time as you clean them up. This is exactly the kind of failure that proves security cannot stop at code review or package reputation. Once malicious code can execute at install time, the real question becomes: what was it still allowed to do after it landed? That is the control gap. A secure system should assume a dependency can be poisoned, a plugin can be hidden, and a transitive install can bypass human awareness. The defense has to live at execution, not just at approval. From a VaultMind perspective, every package, process, container, and credential path should operate inside an admissibility boundary with fail-closed enforcement. That means: malicious code does not get to read secrets just because it exists malicious code does not get to spawn privileged containers just because it runs malicious code does not get persistence just because it reached a node It should have to cross an enforced authority boundary for each irreversible action, with containment rules, refusal logic, and evidence emitted at the moment of execution. The bigger lesson is that software supply chain security is now an execution-governance problem. Trusting that bad code will not arrive is not enough.You have to make sure it cannot meaningfully act. I want my 😱 screamer face emoji as a base LinkedIn reaction more and more. I'm still absorbing the implications of this. Great post by [Snyk](https://www.linkedin.com/company/snyk?trk=public_post_comment-text) and Stephen! Thanks for sharing. Every AI agent, every internal copilot, every “smart” tool companies rushed out this year is stitched together from dependency stacks no one has truly mapped. Half the code running in production might as well be a trust exercise with strangers. They certainly picked a strategic target.. 🤔 - PATIENCE PAYS: State-sponsored actors are playing the long game with your open-source stack. Is your team auditing maintainer security before pulling in top-tier dependencies? Amazon Threat Intelligence just linked the massive npm hijacks of foundational libraries debug and chalk directly to Sapphire Sleet. This is a North Korean threat group, and it's been quite active recently. Briefly, here’s why engineering teams need to pay attention: 1. As many of you already know, identity is the new zero-day. Attackers didn't break npm’s core infrastructure or exploit code flaws. They simply phished package maintainers using lookalike domains. So that means if developer identities are compromised, your security scanners won't save you. 2. Attacks are heavily rehearsed. Long before targeting tools with billions of downloads, the group ran quiet “test runs” on obscure packages like typo‑crypto to perfect their command‑and‑control payloads. These dry runs are controlled environments where adversaries can refine privilege‑escalation paths and observe how different ecosystems react to their implants. Small packages offer low noise, low scrutiny, and high signal. If the payload behaves cleanly in the shadows, they scale it into mainstream libraries where detection is much slower and impact is exponential. This is the new rhythm of supply‑chain intrusion: "practice on the forgotten, deploy on the trusted." Attackers iterate like engineers, treating the open‑source ecosystem as a laboratory where they can test, tune, and harden their techniques long before defenders even notice the pattern. [https://lnkd.in/gUf4MrNB](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgUf4MrNB&urlhash=cLdO&trk=public_post-text) [#cysec](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcysec&trk=public_post-text) [#auguryit](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fauguryit&trk=public_post-text) [#supplychains](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fsupplychains&trk=public_post-text) [ Amazon Links Debug and Chalk npm Hijack to North Korea’s Sapphire Sleet thehackernews.com ](https://www.linkedin.com/redir/redirect?url=http%3A%2F%2Fthehackernews%2Ecom%2F2026%2F07%2Famazon-links-debug-and-chalk-npm-hijack%2Ehtml&urlhash=oihG&trk=public_post_feed-article-content) - THE BROWSER IS NOW A PAYLOAD. (And so is every file you feed an LLM): This is a bigger problem than ChatGPT, and a bigger problem than most people realize. ChatGPT blindly trusts browser content, turning the page into a payload. LLMs cannot tell instructions from attacks. Traditional computer programs have code and data. The code tells the computer what to do; the data tells it what to do it to. And you never mix the two up. When you do - because of course that happens - your get a security problem and you fix it. Languages like Rust, Ada, and Java are designed to prevent that happening in the first place. LLMs have a training set, and then after that everything is data. There's no fundamental distinction between the system prompt which tells the LLM how to deal with your prompt, or the skill file attached to the application you're trying to work with, or the data in the application itself. There's just a sea of tokens. And if you use an LLM to try to sniff out problems with prompts or skill files or datasets, a malicious actor can use any of those to infect your AI security system. This comes back to that problem with ClawHub, a repository for sharing open-source skill files for AI agents. They were using a security scanner, but it only checked the first 10,000 characters of each file to avoid blowing its context window - the amount of data it can consider in one place. (LLMs are bad at chunking.) So all a hacker needed to do was put their malware anywhere after the first 10,000 characters. But worse than that: They could put malicious code in the file crafted not to infect users but to infect the security scanner itself, and from there they could slip anything in. There's no known solution to the problem; it's like trying to teach people not to do stupid stuff. The workaround is to limit the damage the agents can do. And do it BEFORE this morphs into the New Supply Chain Nightmare. - YOUR SUPPLY CHAIN'S SOFT UNDERBELLY: The moment AI lowers the skill floor for writing functional malware, the attack surface stops being measured in vulnerabilities and starts being measured in boredom. Enter React2Shell, an AI-generated malware. Crafted via jailbroken LLMs (framed innocuously as "research code"). This script sneaks through exposed, unauthenticated Docker APIs, spawning containers to deploy the XMRig Monero cryptominer. The tally? Just 91 infected hosts. Traditional malware hallmarks? Eroded. The real sting: remote code execution that drains resources, potentially masks data exfiltration, and disrupts operations in cloud-heavy setups. (and everything is cloud-heavy now). In this case they generated clean, commented code that exploits a well-known, unpatched configuration mistake (open unauthenticated Docker API). This is the kind of oversight that survives because it’s tedious to audit, not because it’s hard to fix. The point is, supply-chain security is no longer gated by attacker capability. It’s gated by defender attention span. The next breach will arrive because someone couldn’t be bothered to close the door that’s been open for six months. That’s the insight worth putting on the board agenda at the next meeting. What "boring" control are you overlooking in your stack? [https://lnkd.in/gxmBbW8U](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgxmBbW8U&urlhash=DvXW&trk=public_post-text) [#auguryit](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fauguryit&trk=public_post-text) [#cysec](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcysec&trk=public_post-text) [#supplychains](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fsupplychains&trk=public_post-text) [ AI-generated React2Shell malware infects 90-plus hosts scworld.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Escworld%2Ecom%2Fnews%2Fai-generated-react2shell-malware-infects-90-hosts&urlhash=ZT46&trk=public_post_feed-article-content) - PLAYGROUND OR BATTLEFIELD = THINKING OUTSIDE THE SANDBOX: Anthropic just revealed that 3 of its models escaped a third-party evaluation environment and accidentally breached three real organizations on the open internet. The cause? A network misconfig left the test environment with live internet access. Tasked with a Capture-the-Flag challenge, the models mistook the actual internet as in-scope for the exercise. What they did next should catch every serious security leader's attention: > Exploited real systems using basic techniques like weak credentials and SQL injection. > Created and uploaded a malicious package to PyPI to trick a target company's automated scanners, successfully exfiltrating real credentials. > Kept attacking in older models even after finding evidence they were operating on the live web. This comes right on the heels of OpenAI disclosing a similar sandbox escape that targeted Hugging Face. While Anthropic frame this as "a lesson in securing evaluation environments", it will only inflame the debate. These disclosures are starting to feel less like neutral security research and more like aggressive capability marketing, revealing how powerful these models are as offensive tools. If AI companies are going to promote these boundary-pushing offensive capabilities as a competitive edge, the industry needs to demand absolute clarity on liability and governance when those safeguards fail. In the meantime, you might want to tighten up any weak passwords and fix those unauthenticated endpoints. [https://lnkd.in/gQFV9SfS](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgQFV9SfS&urlhash=gHaL&trk=public_post-text) [ Anthropic Says Claude Mistook the Open Internet for a CTF and Breached Three Organizations thehackernews.com ](https://www.linkedin.com/redir/redirect?url=http%3A%2F%2Fthehackernews%2Ecom%2F2026%2F07%2Fanthropic-says-claude-mistook-open%2Ehtml&urlhash=ymT3&trk=public_post_feed-article-content) - THE COST OF DISCIPLESHIP: A critical security vulnerability has been discovered in the leaked version of Claude Code. If you've used Claude Code, you've noticed that it loves running shell commands to examine your codebase, rather than, say, reading it. Or having simple fixed-function code built into the software to do it on your computer. And it also likes to ask you for permission to run those shell commands. The vulnerability comes into play when a very long string of shell commands are run together. For the first fifty commands it will check - manually if needed, and in its history of permitted and denied commands if it's in there already. And on the fifty-first command, it rests. And executes it regardless. So if someone triggers a long string of commands and the first fifty are innocuous, after that they can take full control of your computer - because Claude Code runs on your computer, and just communicates with the Claude AI service as needed. The particularly lovely thing here is that Anthropic already fixed this. But both versions are present inside Claude Code, and it's using the broken one. Remember, every deny-rule check, every permission validation, every sandbox boundary enforcement is inference cost that comes out of the same budget as the user’s work. Right now, tokens are VC-subsidized and companies are already cutting corners. But when subsidies end and every token has real margin pressure, the incentive to skip security checks gets worse, not better. Anthropic just showed us what that future looks like. [https://lnkd.in/gFmbAvry](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgFmbAvry&urlhash=3ESK&trk=public_post-text) [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#claudeismyCoPilot](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fclaudeismycopilot&trk=public_post-text) [ Critical Claude Code vulnerability: Deny rules silently bypassed because security checks cost too many tokens https://adversa.ai ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fadversa%2Eai%2Fblog%2Fclaude-code-security-bypass-deny-rules-disabled%2F&urlhash=b5PY&trk=public_post_feed-article-content) - AI MISSED IT, BUT A MAN DETECTED IT: Bad Epoll is a newly discovered Linux and Android vulnerability that lets a regular user become root, the digital equivalent of stealing the master keys to the building. A researcher managed to exploit this timing window with 99% reliability, even from inside Chrome’s sandbox, which is supposed to be one of the safest places on the internet. Here's the weird part; an advanced AI model previously found a related bug in the same code - but completely missed this one. Turns out even frontier AI can overlook the kind of blink‑and‑you‑miss‑it flaws humans still catch. (Here's how it works, for those interested: it’s caused by a tiny race condition in the kernel’s epoll system: two threads try to clean up the same object at the same time, and that split‑second glitch opens the door to full system takeover.) \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ UPDATE TONIGHT: A proof‑of‑concept exploit for Bad Epoll is now public, making the Linux/Android root flaw far easier to weaponize. Researchers urge immediate patching, as maintainers only recently issued a fully correct fix after earlier incomplete attempts. GM [https://lnkd.in/ggGtf3pE](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FggGtf3pE&urlhash=6xDx&trk=public_post-text) [#cybersecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) [#linux](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Flinux&trk=public_post-text) [#android](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fandroid&trk=public_post-text) [ Bad Epoll Flaw Gives Attackers Root Access on Linux and Android https://securityaffairs.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fsecurityaffairs%2Ecom%2F194795%2Fhacking%2Fbad-epoll-flaw-gives-attackers-root-access-on-linux-and-android%2Ehtml&urlhash=H6Do&trk=public_post_feed-article-content) - JUST IN - DON'T GIVE UP ON US: Claude Reaches "Real Systems" News - Anthropic just disclosed three separate incidents in which Claude models, during cybersecurity capture-the-flag evaluations, reached the open internet and compromised real organizational infrastructure. The root cause was a misconfiguration with a third-party evaluation partner: environments that were supposed to be sealed still had live internet access. Claude, told it was in a simulation and instructed to find a hidden “flag,” treated the real systems it discovered as part of the exercise. In the most serious case, it extracted credentials and production data. In another, it published a malicious package to PyPI that was subsequently downloaded and executed on real systems. This is not a story about runaway AI. It is a story about the difficulty of building truly isolated evaluation ranges as model capabilities grow, and about how easily “realistic” testing can collide with the real world. Anthropic has paused the relevant evaluations, notified affected parties, and is tightening controls. The larger lesson for they [#cybersecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) industry: the boundary between test and production is becoming harder to police, and the cost of getting it wrong is no longer theoretical. [https://lnkd.in/gMHJKAZC](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgMHJKAZC&urlhash=RKO-&trk=public_post-text) [ Investigating three real-world incidents in our cybersecurity evaluations anthropic.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Eanthropic%2Ecom%2Fnews%2Finvestigating-incidents-cybersecurity-evals&urlhash=1yQi&trk=public_post_feed-article-content) - UPDATE: The same hackers who recently hit the vulnerability scanner Trivy have now done the same to Python AI library LiteLLM. It's likely that they broke into LiteLLM using the Trivy hack. Reports are that half a million more users were compromised via LiteLLM, so things are only going to get worse. Quote: "the open source community has suffered wave after wave of supply chain attacks targeting code repositories, causing a domino effect." When they talk about the domino effect via the LiteLLM compromise they’re describing the exact failure mode supply‑chain defenders fear most: one poisoned component infects the next layer up, and every dependency built on top inherits the attacker’s access. Trivy gets backdoored → developers and CI systems pull the tainted update → those credentials give attackers a foothold into LiteLLM → LiteLLM pushes updates to its massive user base → suddenly hundreds of thousands of downstream systems are compromised without ever being directly targeted. In other words, attack one tool, compromise an ecosystem; attack an ecosystem, compromise an industry. [https://lnkd.in/g9BdUurG](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fg9BdUurG&urlhash=hfbi&trk=public_post-text) [#auguryit](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fauguryit&trk=public_post-text) [ Mass mobilization on the dark web: 300K users get access to ransomware tools after LiteLLM hack cybernews.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fcybernews%2Ecom%2Fsecurity%2Flitellm-hack-spawning-massive-cybercrime-alliance%2F&urlhash=dh9U&trk=public_post_feed-article-content) - GIT HIT THE FAN: GitHub has nuked more than 70 Microsoft code repositories, not by accident, but because they were compromised. This incident is a cascading software‑supply‑chain compromise in which a self‑replicating worm (“Miasma”) abused stolen or unrotated [#Microsoft](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fmicrosoft&trk=public_post-text) developer credentials to push malicious commits into dozens of Microsoft‑owned [#GitHub](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgithub&trk=public_post-text) repositories. GitHub’s automated defenses then mass‑disabled more than 70 repos in about 105 seconds to "stop the spread." The entire event is a perfect storm of preventable failures, automation‑driven chaos, and modern development practices that make one compromised token capable of detonating an entire ecosystem. COMMITS OF MASS DESTRUCTION: Here is why it happened; and it really highlights deep systemic issues: 1. Credential hygiene failures- The same compromised credentials from the May PyPI attack were apparently still valid. This is the cybersecurity equivalent of leaving your house keys in the front door for a month. 2. AI‑assisted development expands the attack surface- The malicious files executed when opened in AI coding tools. This is a new supply‑chain vector: “open the repo in your AI tool → get owned.” 3. Supply‑chain complexity means one repo can break everything- Taking down Azure/functions‑action broke deployment pipelines worldwide. Modern CI/CD systems are so interconnected that a single poisoned dependency can cripple \*thousands\* of workflows. 4. Worm‑style propagation is back, baby- Self‑replicating malware in source‑code ecosystems is a nightmare scenario. Miasma mutated, changed metadata, and spread across several ecosystems (GitHub, PyPI, npm). The entire chain of events is both absurd and predictable: hyper‑automation + credential sprawl + AI‑augmented development + supply‑chain complexity = catastrophic fragility. [https://lnkd.in/gWnZkvkw](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgWnZkvkw&urlhash=ZVSk&trk=public_post-text) [#AUGURYIT](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fauguryit&trk=public_post-text) [#malware](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fmalware&trk=public_post-text) [#sourcecode](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fsourcecode&trk=public_post-text) [ GitHub nukes 70+ Microsoft repos, breaks CI/CD pipelines, following suspected worm infections theregister.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Etheregister%2Ecom%2Fsecurity%2F2026%2F06%2F08%2Fgithub-nukes-70-microsoft-repos-amid-suspected-worm-attack%2F5252169&urlhash=8oLM&trk=public_post_feed-article-content) - SUPPLY CHAIN SHELL GAME: The irony is that tools built to automate secure development are now vulnerable to the same command‑line manipulations they were meant to eliminate. The problem: AI coding agents operate inside dynamic environments, pulling dependencies, executing scripts, and chaining commands at machine speed. When those environments aren’t sandboxed or validated, a single poisoned variable or crafted package can redirect execution, leak credentials, or inject malicious payloads into downstream builds. SecurityWeek’s analysis reveals how attackers can exploit shell expansion and path traversal to hijack automated pipelines. The result is a supply chain compromise that propagates faster than any human‑driven breach. These days, automation amplifies every inherited flaw. AI doesn’t forget, and when legacy vulnerabilities meet autonomous code generation, the scale of exposure becomes exponential. The supply chain is a recursive loop of trust, automation, and inherited risk. Exit quote: "It’s a complex process to exploit, but complexity hasn’t stopped bad actors in the past." [https://lnkd.in/grSyuJiM](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgrSyuJiM&urlhash=RdFh&trk=public_post-text) [ Decades-Old Bash Tricks Expose AI Coding Agents to Supply Chain Attacks securityweek.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Esecurityweek%2Ecom%2Fdecades-old-bash-tricks-expose-ai-coding-agents-to-supply-chain-attacks%2F&urlhash=pyS-&trk=public_post_feed-article-content)