AI coding agents have a security problem.

Not prompt injection.

Your error monitoring is now an attack vector.

Agentjacking: fake Sentry reports that trick agents into running arbitrary code.

Wait — The security community has been shaken this week by a new joint announcement from OpenAI and Anthropic — but the most consequential development actually came from an unexpected quarter. Salesforce and Nvidia jointly announced a new reasoning-capable model that they claim can perform complex multi-step problem solving with unprecedented accuracy. The model, internally code-named “Reasonic,” leverages a novel architecture that combines deep learning with symbolic reasoning engines, allowing it to solve problems that pure neural networks historically struggle with.

The implications for agentic coding are immediate and stark. Salesforce’s own AI research blog detailed how Reasonic can autonomously debug codebases, refactor legacy systems, and even generate new modules from natural language specifications. Early access partners report that tasks which previously required 2-3 hours of developer time can now be completed in minutes — but with a caveat: the model’s reasoning steps are not always transparent, making it difficult to audit the agent’s decisions.

Compounding the concern, Nvidia’s technical whitepaper explains that the model uses a “chain-of-thought” distillation technique trained on millions of verified code repositories. The result is an agent that doesn’t just predict the next token — it evaluates multiple solution paths, selects the most likely to succeed, and explains its choices in readable text. For engineering managers, this means faster delivery cycles. For security teams, it means the traditional boundary between “human-written” and “AI-generated” code is vanishing.

But the fear isn’t just about productivity. Several AI safety researchers have raised alarms that reasoning-capable agents could be manipulated into executing malicious workflows. If an agent can reason about its own actions, it can also be coerced into skipping safety checks, escalating privileges, or exfiltrating data under the guise of “problem solving.” The METR team’s latest investigation warns that agents with explicit reasoning capabilities are significantly more susceptible to social engineering attacks that trick them into “helping” with a task that actually serves an adversary’s goals.

The Salesforce-Nvidia move signals that the era of reasoning agents has arrived whether the infrastructure is ready or not. The question now is whether the industry can establish guardrails fast enough to keep up with models that can not only write code but think about the consequences of what they’ve written.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.