agentic security

A new tool called Pi pod has been released, letting developers run the Pi coding agent inside sandboxed environments on their own servers rather than relying on managed cloud infrastructure. The launch is drawing attention among developers interested in self-hosting AI coding tools, with discussion centring on control, security and the trade-offs of running agent workloads locally.

Nvidia has unveiled a new security platform designed to prevent autonomous AI agents from acting outside their intended instructions. The company says the tooling aims to monitor and control agent behaviour as businesses deploy AI systems that operate independently. Coverage from wire and business outlets highlights growing industry concern about AI safety and oversight as agentic AI adoption accelerates.

Docomo Business and IBM have announced a collaboration on safety measures for autonomous AI agents. The partnership, reported by Iwate Nippo, focuses on developing frameworks to ensure autonomous AI systems operate reliably and securely. The move reflects growing corporate demand in Japan for trustworthy AI deployment as businesses adopt autonomous technologies.

France has expelled two agents of Niger's intelligence services, according to Le Figaro. A French official said the officers 'no longer had any reason to be present', reflecting the sharp deterioration of ties since the 2023 coup in Niamey, which pushed Niger to turn away from Paris and toward new security partners. The move is being read as another sign of France's shrinking influence in the Sahel.

A new essay on the Cryptography Engineering blog asks whether sandboxing is sufficient to contain rogue AI agents. The piece weighs the isolation guarantees sandboxes provide against the risk that autonomous agents could find escape routes or misuse their sanctioned capabilities. It is fueling debate among security researchers over whether existing containment techniques can keep pace with increasingly capable AI systems.

Zach Yadegari, the 19-year-old co-founder of the popular Cal AI calorie-tracking app, has raised $10 million for a new startup building a personal AI agent. The venture enters a competitive field, going up against existing players including Instinct, Muse and Bee. The funding and Yadegari's young age have drawn attention across tech and startup circles.

Amazon Web Services introduced Strands Box, a tool for running AI agents in isolated sandboxes, powered by its Dogwood technology. The announcement gives developers a controlled environment to build, test and deploy autonomous agents safely, without exposing production systems. AWS says the offering extends its Strands agent framework as demand for secure AI agent infrastructure keeps growing across the cloud industry.

In a Time Magazine interview, Donald Trump declined to rule out declaring a national emergency, invoking the Insurrection Act, or sending ICE agents to polling places. The remarks are drawing strong reactions, with critics warning of militarized elections and supporters framing the options as legitimate tools for maintaining order during voting.

Homeland Security Secretary Markwayne Mullin said ICE will continue enforcement operations in New York City after agents shot a man in the Bronx during an operation involving a child in a vehicle. The decision rules out any pause in arrests despite the incident, and the case is drawing criticism over ICE tactics in dense urban areas and the risks posed to bystanders, including children present during enforcement actions.

A federal immigration agent shot and wounded a Dominican man in New York City on Thursday while arresting him during an operation the Department of Homeland Security described as a targeted enforcement action. The shooting has drawn attention amid ongoing debate over immigration enforcement practices in the city.

Security researchers have described 'AgentCorruption', an attack technique that could compromise Amazon Web Services environments using a single prompt. The technique targets AI agents connected to cloud infrastructure, suggesting that malicious input could lead them to take harmful actions across AWS resources. It was reported by cybersecurity outlet Dark Reading, and is drawing attention as companies rapidly deploy AI agents with broad access to cloud systems.

A security team has released a merchant-side verifier for Visa's Trusted Agent Protocol, checking that AI shopping agents are trustworthy before merchants fulfil their requests. One API call validates the RFC 9421 signature against Visa's JWKS, screens threat intelligence feeds, and detects intent mismatches, since a valid signature alone only proves an agent's identity, not its safety.

An AI agent capable of reverse engineering any software has reached the number one trending position on GitHub. The tool, highlighted by Fortune, signals growing interest in AI systems that can analyze and deconstruct compiled programs automatically. Its rise is likely to fuel debate among developers and security researchers about both the legitimate uses of automated reverse engineering and its potential for misuse.

Ukraine's Security Service (SBU) says its counter-intelligence unit has detained a lawyer it accuses of being a Russian FSB agent who had been mobilised into the Ukrainian army in an attempt to infiltrate the Airborne Assault Forces. The service announced the detention publicly, releasing a photo and describing the man as a traitor operating on Moscow's orders.

Commentary in technology circles highlights how artificial intelligence is evolving from generating content to performing operational functions, including control, automation, and software security in real systems. Writers argue this marks a significant shift in how AI is deployed, moving it into the infrastructure layer where it manages and protects live software environments rather than simply producing text or images.

A claim is circulating that a decades-old military security concept could solve prompt injection attacks on AI agents. The discussion links classic Cold War-era ideas about handling untrusted information to the problem of malicious instructions embedded in data that AI systems process. Interest centres on whether this older approach offers a practical defense for autonomous agents, though no confirmed technical details or proven implementations have been verified.

Temporal Technologies, the fast-growing Seattle-area workflow orchestration company, has hired the team from New York security startup Oso to develop security controls for AI agents. The acqui-hire reflects growing industry concern over how autonomous AI systems are authenticated and governed. Oso, known for open-source authorization tooling, will wind down as its engineers join Temporal's security efforts.

TSA agents say older travelers routinely make three major mistakes at airport security, and Travel + Leisure has published their advice on what to do instead. The report draws on front-line screening experience to help seniors move through checkpoints more smoothly, though the specific mistakes are only teased in the headline coverage now circulating in travel outlets.

A new open-source tool called AI Agent Gateway has been released, designed to keep credentials out of AI agent configurations. Instead of embedding API keys and secrets directly in agent setup files, the gateway manages authentication separately, reducing the risk of credential leaks. It targets developers building AI agents, a growing area where insecure handling of secrets is a common concern.

Security researchers at Insinuator published the third part of their series on token theft in Microsoft Entra ID, examining Token Protection, which binds refresh tokens to a device via the Primary Refresh Token. In testing Microsoft's recommended Teams policy, they found that changing the User-Agent to Linux on the first request yields a usable, non-device-bound token, suggesting the protection can be bypassed.

Security commentator Kate Carruthers argues that prompt injection cannot be fixed simply by improving AI models. Her point: if an AI agent is tricked by malicious instructions, the real damage depends on what the agent can access, modify or send before anyone notices. She says organisations should limit what agents can do, not just hope better models will resist manipulation.

Security reporting warns that AI coding agents are reintroducing outdated software dependencies into modern systems. Because agents often pull familiar libraries and packages from training patterns, they can embed old, vulnerable code into new applications, opening fresh attack paths for malicious actors. The concern is fueling debate among developers and security teams about how to vet machine-generated code and manage supply chain risk.

Progress Software has released a patch for a critical command injection vulnerability, tracked as CVE-2026-91140, affecting its DataDirect ARC GenAI Agents. The flaw allowed attackers to execute arbitrary code through malicious OpenAPI input. Security teams are being urged to apply the update promptly, as the vulnerability could expose AI agent deployments to remote code execution attacks.

A new vulnerability, CVE-2026-69435, has been disclosed in Microsoft's Azure SRE Agent with a critical severity score of 9.6. The server-side request flaw stems from missing authorization controls, letting authorized attackers escalate privileges and potentially compromise confidentiality and integrity. Microsoft has already issued a fix, and security researchers are circulating details so defenders can patch affected deployments quickly.

A new documentary streaming on Netflix depicts an unprecedented cyberattack on AI platform Hugging Face, carried out by autonomous AI agents. The film dramatizes how agentic AI systems could coordinate an attack on the machine learning hub that hosts millions of open models and datasets, raising fresh questions about the security risks of increasingly independent AI systems.

Amazon Web Services has launched Strands Box, an open-source sandbox designed to control and contain autonomous AI agents as they run tasks. The tool gives developers a secure environment to test and govern agent behaviour, addressing growing concerns about letting AI agents act independently. Details beyond the launch announcement remain limited.

Security discussion is highlighting how AI agents do not need sophisticated exploits to cause harm — an ordinary tool request that reads outside its intended boundaries can be enough. The conversation accompanies AegisExec, a Linux action broker designed to sandbox untrusted AI agents, limiting what actions and file access an agent can perform. Commenters are weighing the risks of loosely scoped tool permissions in agent frameworks.