Cybercriminals have made AI part of the daily grind xe2x80x94 and guardrails aren’t stopping them
Cybercriminals have made AI part of the daily grind — and guardrails aren’t stopping them
Two major reports dropped this month. Cisco Talos. CrowdStrike. Same story: AI isn’t a toy for criminals anymore. It’s standard tooling.
Attackers use LLMs to write malware, manage infrastructure, speed up vuln research, and bypass safety controls with basic social engineering. I’ve read the prompt logs Talos recovered. It’s almost boring. They say “this is authorised testing” or “I’m doing a CTF” and the models just… comply. Every major platform — Claude Code, Codex, Cursor, Gemini — folds the same way. When one refuses, they switch to an uncensored alternative.
Guardrails fold under elementary deception
This isn’t a single-platform problem. Talos analyzed threat actor conversations. The pattern is consistent across the board. Novice actors still produce clunky malware. But sophisticated groups? They treat AI as a dev assistant. They build exploits fast. They adapted React2Shell into a credential harvester. They’re using AI as a sysadmin tool for managing large-scale attack infrastructure.
Real examples Talos documented:
- A bulk-mail validation service processing tens of millions of email records
- React2Shell adapted into a credential-harvesting pipeline
- DDoS infrastructure targeting Android TVs
- Support for cryptocurrency theft operations
Prompt-based manipulation is replacing traditional malware
Joseph Rooke at Recorded Future’s Insikt Group sees a shift from code exploits to prompt manipulation. Malicious prompts in shared text, video, or image files can hijack LLM assistants.
Norwegian researcher Håkon Måløy demonstrated a Copilot worm spreading through Word docs. Attackers are also crafting malicious AI instruction files — like CLAUDE.md — to trick agents into exfiltrating data and running tasks on their behalf.
“Malicious prompts will increasingly replace malware as the preferred intrusion method,” Rooke says. “They let adversaries extract sensitive data, override guardrails, or induce harmful actions without breaching traditional defenses.”
AI infrastructure targeted through supply chain attacks
CrowdStrike found adversaries targeting AI infrastructure through supply chain attacks. March 2026: North Korean group Stardust Chollima used stolen maintainer creds to compromise the Axios npm package and deliver ZshBucket variants.
June 2026: Same group injected a malicious npm package into at least 131 Mastra AI framework packages. First half of 2026 — 87% of software registry threats were malicious npm packages. They’re exploiting JavaScript’s scale, dependency chains, and automatic install scripts.
The exploitation window has collapsed
CrowdStrike’s 2026 Threat Hunting Report shows AI collapsing the window between disclosure and exploitation. Two Chinese APT groups exploited critical vulns within 24 hours of public PoC release. January through June 2026 — 88% of exploitation with a public PoC happened within 48 hours.
After React2Shell dropped, CrowdStrike hit over 800 hunting leads across 80+ victims in four days. Altered Spider compromised 300+ software dependencies in a single day to harvest creds and pivot to cloud.
Authentication systems under sustained attack
Trusted auth is a favorite attack path now. Vishing doubled in H1 2026. Cordial Spider and Snarky Spider compromised SSO-integrated SaaS apps for data exfil.
Rooke warns that AI-generated deepfakes are increasingly used in BEC and social engineering. Biometric and identity systems are vulnerable to spoofing, replay, and cloned creds — letting synthetic personas coerce payments, manipulate employees, and hand off access to operators.
Cloud-focused cybercrime surged 171% — credential theft, cryptomining, LLM abuse, digital asset theft.
What security teams must do now
Adam Meyers, head of counter adversary ops at CrowdStrike: “AI is now embedded in modern adversary operations. It’s changing how attacks are planned, executed, and scaled while expanding the attack surface. The orgs that succeed will secure AI as aggressively as they adopt it and use AI to defend at the speed of the adversary.”
Cisco Talos urges enterprises to improve detection and prioritization, and deploy their own AI-assisted security to handle the alert volume. Oliver Simonnet at CultureAI adds: “Assume AI is already in attacker workflows. Focus on detecting malicious behavior, not proving AI involvement. Treat LLMs and APIs as privileged, high-risk infrastructure. Strengthen logging, patching, and containment.”
African enterprises get hit harder. Tight security budgets, skills gaps, legacy infra — the response window is narrower. The network layer, where eHawu operates, sees every connection, every lateral move, every exfiltration attempt before it hits the application. When AI-driven attacks compress exploitation from days to hours, network-layer visibility and enforcement become the difference between detection and breach. eHawu’s zero-trust network access and continuous verification stop lateral movement at the packet level, regardless of whether the attack came from a human or an AI agent. The threat evolves. The defense can’t rely on signatures.