Artificial intelligence news.
Straight from the original announcements.

Five Security Lessons for AI Agents From All Day AI 2026

November 19, 2026, 20:52By the AI News Bites Editorial TeamSource: Anaconda
Five Security Lessons for AI Agents From All Day AI 2026
Image: Anaconda

Security professionals at the All Day AI Conference 2026 called for hard limits around AI agents, with controls enforced outside the models themselves. Their recommendations cover identity, sandboxing, runtime verification, supply-chain visibility and vulnerability prioritization.

Key points
  • Enforce agent limits outside the model and its control.
  • Use isolated sandboxes, scoped credentials and default-deny egress.
  • Sign and hash artifacts, including model weights, at runtime.
  • Track models, data, plugins and MCP servers as supply-chain components.
  • Reduce shipped software before prioritizing vulnerabilities.

Security boundaries should not depend on the agent

The security talks took place on October 22 at All Day AI, described by Anaconda as a vendor-neutral online conference run by teammates at Anaconda and Kilo under the direction of Brian Turcotte, a senior developer relations engineer at Anaconda.

Chris Hughes, founder of Resilient Cyber, argued in the keynote that alignment and security address different problems. “Alignment (…) is lowering how often the agents go wrong, but security is capping what it costs. And we need both,” Hughes said.

He also questioned the value of routine human approvals. Citing a figure he attributed to Anthropic, Hughes said users approve 93% of Claude Code permission prompts. He described users as a bottleneck when they cannot keep up with agent activity, and said refusals and classifiers are “essentially speed bumps.”

His proposed controls include a unique identity for each agent, credentials injected only when needed, policy written as code, restricted outbound network access and a sandbox. The key principle is that these boundaries should be enforced by the software around the model, rather than relying on the model to follow instructions.

Sandboxing and runtime checks limit attack paths

Madhu Akula, an independent security engineer, treated agents as untrusted workloads and applied container-security practices including least privilege, kernel-level isolation and outbound network rules. He argued that file access, egress, credentials and monitoring should be controlled outside the sandbox, because an agent that can modify a control may weaken or disable it.

In a live demonstration, a malicious dependency stole SSH keys from a developer laptop, sent them to a command-and-control server, collected API keys from the environment and followed instructions in a poisoned README to post credentials and delete files. Inside an isolated sandbox with a credential broker and default-deny egress, the attack chain was blocked at each stage.

Akula also showed how an agent could use a server-side request forgery bug in an artifact registry proxy to reach the internet. The example underscored Anaconda’s broader point that package sources are part of an agent’s attack surface.

RanjithKumar Ragavan of T-Mobile described a different layer of defense. T-Mobile checks feature modules and on-device AI models before they run on 53 million devices. Files are signed at the source, their SHA-256 hashes are registered in a trusted registry and the hashes are checked each time the files load. Failed checks prevent execution, while a server-side kill switch and automatic rollback provide a response when a signed artifact later proves unsafe.

The AI supply chain needs inventory and prioritization

Hasan Yasar of the Software Engineering Institute at Carnegie Mellon University framed AI systems as software supply chains. He identified four tiers: foundation models and pre-trained weights, data and internal retrieval stores, agent skills, plugins and Model Context Protocol servers, and the execution runtime.

His DevSecOps pipeline includes data sanitization, rejection of unsafe file formats such as Python pickle, automated red-teaming, and hashing model weights while generating an AI bill of materials. Yasar said that inventory can include model-weight digests, dataset provenance, system-prompt hashes and MCP server identities. “If you don’t know what your systems are, you cannot build any of this, so know your system first,” he said.

Russ Andersson recommended focusing patching on risk and using Security Technical Implementation Guide scanning to check configurations. His approach starts by reducing the amount of software that needs protection: scan production, remove unused components, and rank the remainder by exposure and evidence of exploitation rather than severity alone.

For users and platform teams, the message is a practical checklist. Give agents scoped identities and credentials, keep sandbox and network controls outside their reach, verify every artifact at runtime, fail closed when checks fail, track skills and MCP servers as dependencies, and reduce what is shipped before prioritizing what remains. Anaconda says its AI Security & Guardrails offering, through Enkrypt AI by Anaconda, supports model, agent and MCP testing, runtime guardrails, access policies and compliance evidence alongside the identity, sandbox and network controls operated by customer teams.

Related news

Anaconda: Two LiteLLM Supply Chain Attacks Never Reached Its Curated Channel
Safety and ethicsNovember 20, 2026, 02:41

Anaconda: Two LiteLLM Supply Chain Attacks Never Reached Its Curated Channel

Anaconda says neither of the two compromised LiteLLM releases made it into its curated Python package channels, crediting a build delay and source-based builds. The…

GitLab warns Claude Code settings file can expose secrets
Safety and ethicsNovember 11, 2026

GitLab warns Claude Code settings file can expose secrets

GitLab has published a guide for finding and removing secrets that may be exposed through Claude Code’s .claude/settings.local.json file. The company says the file can…

Claude Code Settings Exposed 1,379 Secrets in Public GitHub Files
Safety and ethicsNovember 10, 2026, 16:59

Claude Code Settings Exposed 1,379 Secrets in Public GitHub Files

Claude Code’s approval settings exposed 1,379 distinct secrets across 1,122 public GitHub files, according to security company Apiiro’s research reported by The Hacker…