RSAC 2026 in Hindsight: AI Assistants Invented 19.7% of Recommended Packages
The Positive Technologies team re-read the talks from RSAC 2026, a major cybersecurity conference, and checked their predictions. For example, an AI assistant recommends a library that does not exist, and an attacker registers that name with malware inside. In one study, 19.7% of 2.23 million such recommendations were invented. Big vendor numbers without a methodology are only a reason to verify, not to buy. Before installing a package an AI suggests, check its origin and its code.
AI-processed from Habr AI; edited by Hamidun News
In March, RSAC 2026, one of the largest cybersecurity conferences, ran more than 570 sessions, and AI agents took up a noticeable share of the program. The ML team at Positive Technologies picked ten talks and looked at what happened to their predictions after the conference. The main problem is that the headline numbers in the talks are rarely backed by a methodology.
Here is a simple example: you ask an AI assistant to pick a Python library, it names a package that never existed, and an attacker has already registered that name in the registry and put malicious code there. This attack is called slopsquatting, and in one study, 19.7% of 2.
23 million package recommendations turned out to be invented. The practical takeaway: check the origin of your dependencies, and test other people's metrics on your own data. Below are eight talks from the review: what the authors said and where a caveat is needed.
Risks for people who use AI and write code with it
AIKatz (Steve Kupchik, Lumia Security). If a trojan is already sitting on a computer, it can read a token from the memory of a desktop AI app and access the service as the user. The researchers extracted tokens from ChatGPT and Copilot, and a session key from a cookie in Claude.
With a live session, an attacker can simply request the chat history, and through the memory feature can try to influence future answers. Microsoft, Anthropic and OpenAI closed the reports as informational: by their logic, accessing process memory within the same session does not cross a security boundary. The defense is to restrict access to the memory of AI app processes and monitor attempts to read it, but this does not replace workstation protection.
Slopsquatting (Joe Spracklen, USENIX Security 2025). The authors tested 16 models on Python and JavaScript tasks. Of 2.
23 million recommendations, 19.7% were nonexistent packages: at least 5.2% on average for commercial models and 21.
7% for open ones. Repeatability is the most dangerous part: in 43% of cases, the model invented the same name ten times out of ten, so an attacker does not have to guess which name to register. RAG (giving the model a list of known packages) lowered the share of invented names, for example for CodeLlama 7B from 26.
28% to 13.40%. Fine-tuning brought it to 2.
66% and 10.27% in two models, but task-solving in HumanEval, a code-writing benchmark, dropped: for DeepSeek from 51.4% to 25.
3%. This is 2025 data on specific models and cannot be carried over to current tools. The conclusion: check the origin and contents of a dependency, not just whether the name exists.
"Is AI code less secure than human code?" (OX Security). The authors studied more than 300 open-source applications, 50+ of which were vibe-coded, meaning written almost entirely by AI, and asked Cursor, Claude and Codex to build the same service. They found open redirect, path traversal and XSS. But there is no comparison of error rates with human-written code, and the method for checking the repositories is not disclosed, so the talk does not answer the question in its title. What it does show is that generation speed increases the amount of code that needs review.
Attackers and their automation
VibeCrime (TrendAI). AI is being built into the market for crime-as-a-service: phishing, code, processing of stolen data. Many "criminal LLMs" are ordinary models without restrictions or resold access to someone else's API; WormGPT was already being advertised in 2023. The authors call the next step cybercrime-as-a-sidekick: an orchestrator hands out tasks to agents for reconnaissance, finding an entry point and preparing tools. This is a forecast, not a confirmed practice; in the authors' view, economics will determine how fast the shift happens.
Always-on purple team (NVISO and SANS). A purple team runs coordinated exercises in which attackers and defenders look together at what the defenses noticed. The talk's idea: an agent looks for an entry point, and an operator approves the plan. In the demo, an agent on GPT-5 working through OpenClaw and Telegram, with Kali Linux and theHarvester, carried out OSINT reconnaissance of a test target and fixed a problem with an MCP server. They did not show a full chain through to access, and rules for the agent itself, such as who confirms its actions and how its scope is limited, are lacking.
Defenders and their metrics
AI Agents Never Sleep (Erez Harush, Wiz). The agent triages alerts in a SOC (security operations center): the claims are 3000 investigations per month per team, 3–20 minutes per verdict, and $0.5–1.5 per case. This is vendor data without independent verification. The engineering lessons are more useful: with ~30 tools, the agent got confused, and routing examples raised tool-selection accuracy from 60% to 95%; by the eighth step, the agent forgot the original alert, so context compression, subagents and checkpoints are needed. Prompts should be versioned and tested like code.
We Put AI in Our SOC (Gupta and Mittal). Instead of an autonomous analyst, the team built an assistant inside the incident card with read-only access and links to evidence. Time to triage (initial assessment) fell from 42 to 27 minutes, and time to case closure from 3.2 to 2.5 hours. The project ran for six months and is at the evaluation stage; problems such as prompt injections in logs were handled by normalizing data and separating untrusted text from instructions.
SOC detectors (Dorie Fischer, Swordfish). According to data from a large MSSP (managed security service provider, 82 620 records), 93.4% of ML rules and 94.7% of static rules produced not a single confirmed threat in a year. But that does not yet mean the rules are useless: they may catch rare events or duplicate other rules. Fischer proposes counting not alerts but incidents: whether the incident was found, how quickly the first useful signal arrived, and what the detector added to coverage.
What follows from this
The general lesson: forecasts and vendor numbers without a methodology are a reason to verify, not to buy. For developers, that means controlling dependencies and reviewing generated code. For SOC teams and founders, it means measuring the effect on your own data, limiting the agent's permissions, and separately testing integrations, where a "hallucination" often turns out to be a broken connector.
- Nov 19, 2026Codex Agents Ran 1700 Investigations Into ChatGPT Slowdowns and Found the Cause in 600
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Infostart connects Direct, Metrica and Wordstat to ChatGPT: 6 commands, shared access
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
Key Questions Answered
What are the key details on Risks for people who use AI and write code with it?
AIKatz (Steve Kupchik, Lumia Security). If a trojan is already sitting on a computer, it can read a token from the memory of a desktop AI app and access the service as the user. The researchers extracted tokens from ChatGPT and Copilot, and a session key from a cookie in Claude. With a live session, an attacker can simply request the chat history, and through the memory feature can try to influence future answers.
What are the key details on Attackers and their automation?
VibeCrime (TrendAI). AI is being built into the market for crime-as-a-service: phishing, code, processing of stolen data. Many "criminal LLMs" are ordinary models without restrictions or resold access to someone else's API; WormGPT was already being advertised in 2023. The authors call the next step cybercrime-as-a-sidekick: an orchestrator hands out tasks to agents for reconnaissance, finding an entry point and preparing tools.
What are the key details on Defenders and their metrics?
AI Agents Never Sleep (Erez Harush, Wiz). The agent triages alerts in a SOC (security operations center): the claims are 3000 investigations per month per team, 3–20 minutes per verdict, and $0.5–1.5 per case. This is vendor data without independent verification.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.