Showing 56 dispatches
Untrusted text reaching an LLM is the injection flaw of this decade. Here is how prompt injection works, why it is so hard to fix, and what actually reduces the blast radius.
Read dispatch
Retrieval-augmented generation pulls in documents you don't fully control. Every one of them is an instruction the model might obey.
Read dispatch
If an attacker can slip crafted samples into your training data, they can plant behaviour that survives into production. Poisoning is a supply-chain problem for models.
Read dispatch
A trained model is a lossy compression of its data. With the right queries, attackers can reconstruct fragments of it or prove someone was in the set.
Read dispatch
OWASP mapped the ten risks that matter most for large language model apps. Here is a practitioner's tour of each, with what to actually do about it.
Read dispatch
Guardrails are the seatbelts of LLM apps. Done well they catch harm; done badly they break the product and still miss attacks.
Read dispatch
The moment an LLM can call tools, send emails or move money, a prompt becomes a command. Agentic AI turns content risks into action risks.
Read dispatch
You didn't train it, you downloaded it. Pre-trained models, datasets and pip packages are a trust chain most teams never audit.
Read dispatch
A few carefully placed pixels can make a model call a stop sign a speed limit. Adversarial examples are perturbations invisible to us but decisive to the model.
Read dispatch
AI apps sprinkle API keys, vector DB creds and model tokens everywhere. Prompts and logs then quietly exfiltrate them.
Read dispatch
Every model ships with rules and every week someone talks it out of them. Jailbreaks reveal the gap between alignment and enforcement.
Read dispatch
Embeddings feel anonymous. They aren't. Given the vectors, an attacker can often reconstruct the text they came from.
Read dispatch
You threat-model your web app. Your model deserves the same. ML systems have attack surfaces classic threat models never enumerate.
Read dispatch
Your staff are already pasting source code and customer data into chatbots you never approved. Shadow AI is the new shadow IT.
Read dispatch
Red teaming an LLM is not the same as pentesting a web app. Here is how to think like an adversary against a model that talks back.
Read dispatch
Ad-hoc poking finds ad-hoc bugs. A repeatable methodology finds the ones that matter, every time, and proves coverage.
Read dispatch
Fuzzing broke C programs for decades. Pointed at models, it surfaces the weird inputs that make them crash, leak or misbehave.
Read dispatch
The system prompt is supposed to be invisible. Often it is one clever question away from full disclosure - along with its secrets.
Read dispatch
One tester and a coffee will not cover a guardrail's failure surface. Automated attack suites test thousands of variations while you sleep.
Read dispatch
Give an agent tools and you give an attacker a toolbox. Testing agentic AI means testing what it can be tricked into doing.
Read dispatchA biased model is not just an ethics problem - it is an exploitable one. Fairness gaps become attack surface and liability.
Read dispatchWhen a model can read images, an attacker can hide instructions inside a picture. Multimodal expands the injection surface dramatically.
Read dispatchA confident wrong answer can be worse than no answer. To manage hallucination you first have to measure it rigorously.
Read dispatchYour RAG app trusts its documents. Plant one malicious paragraph in the knowledge base and the model works for the attacker.
Read dispatchYou don't have to build every probe from scratch. Open tooling like garak and PyRIT industrialise LLM security testing.
Read dispatchA model update can silently reopen a bug you closed months ago. Without regression tests, safety quietly rots.
Read dispatchA real AI pentest is scoped, threat-led and evidence-driven - not a weekend of jailbreak screenshots.
Read dispatchYou can't DDoS a serverless AI app off the internet - but you can bankrupt it. Denial-of-wallet is model DoS with a billing twist.
Read dispatchScanners produce noise; humans burn out. LLMs can pre-triage findings so analysts spend attention where it counts.
Read dispatchAI reviewers never get tired at line 4,000. Used well they widen coverage; used blindly they rubber-stamp vulnerabilities.
Read dispatchSignatures catch what you've seen. Anomaly detection aims at what you haven't - the price is false positives you must engineer around.
Read dispatchAI can read the whole attack surface faster than any analyst. That power cuts both ways, so it comes with rules of engagement.
Read dispatchA SOC drowns in alerts. LLMs can summarise and correlate them - as long as the summary never hides the one that matters.
Read dispatchBlocklists are always a step behind. Language models can read intent, catching the well-crafted lure a static filter waves through.
Read dispatchDumb fuzzers stumble on bugs by luck. AI-guided input generation reaches deep code paths a random mutator never would.
Read dispatchNo team can read every advisory, forum post and CVE. AI can - and can hand analysts the 1% that matters to them.
Read dispatchLogs hold the evidence and the noise in equal measure. AI can cluster, baseline and narrate them so investigations start hours ahead.
Read dispatchAn AI copilot can draft your payload, explain a protocol and waste an hour on a confident wrong turn. Knowing which is the skill.
Read dispatchPolymorphic malware defeats hashes by design. ML classifiers learn the shape of malice - and attackers learn to reshape it.
Read dispatchStatic honeypots fool no one for long. LLM-backed decoys can hold a conversation, keeping an intruder busy and observed.
Read dispatchDescribe the behaviour in plain English; get a draft detection back. AI is shrinking the gap between an idea and a working rule.
Read dispatchThe goal isn't to replace the analyst - it's to give them a tireless junior who reads everything and forgets nothing.
Read dispatchZero trust is not a product you buy - it is a principle: never trust, always verify. Here is what that means in practice.
Read dispatchPasswords fall daily to phishing and reuse. MFA is the single highest-leverage control most organisations still under-deploy.
Read dispatchRansomware rarely detonates on day one. It is the loud finale of a quiet intrusion you could have caught days earlier.
Read dispatchThe password isn't dead yet, but passkeys are writing the obituary. Here is how to bridge the two worlds safely.
Read dispatchATT&CK is the shared language of how attackers actually operate. Learn to read it and your detections stop being guesswork.
Read dispatchNo single control holds forever. Defence in depth assumes each layer will fail and makes sure the next one catches it.
Read dispatchThe easiest way past a firewall is to ask someone to open the door. Technology can't patch a helpful human.
Read dispatchMost cloud breaches aren't clever exploits - they're an open bucket and a wildcard IAM policy nobody reviewed.
Read dispatchHow you spend the first hour of an incident decides how the next month goes. Panic is expensive; a plan is priceless.
Read dispatchYou don't just run your own code - you run everyone's. A single poisoned dependency can undo every other control.
Read dispatchYour APIs do the real work and hold the real data - and they're often protected by little more than optimism.
Read dispatchSeeing is no longer believing. A cloned voice or a fabricated video call is now a viable path to a fraudulent wire transfer.
Read dispatchTools don't secure a company; people do. Culture is the control that scales when your headcount outpaces your policies.
Read dispatchFinding vulnerabilities is easy; fixing the right ones fast is the hard, unglamorous work that actually reduces risk.
Read dispatch//: SIGNAL BOOST
Follow Vantixia on LinkedIn and our WhatsApp channel for new dispatches as they drop - or bring these problems to our team directly.
