Help Net Security · Oct 13
NinjaOne launched Katana AI and an MCP server that puts outside AI agents under the same permissions as technicians
NinjaOne announced Katana AI, an agentic technician inside its IT operations platform that reads device telemetry, history and alerts to suggest likely root causes, remediation steps and scripts. Alongside it, the NinjaOne MCP server lets external AI agents read NinjaOne data and take actions, governed by the role-based permissions IT teams already manage. NinjaOne did not give availability dates or pricing.
▸ The MSP Angle
Can I let AI agents take actions inside my RMM?
Increasingly yes, but only through a controlled door. An MCP server that applies existing role permissions means an agent can be scoped like a junior technician instead of handed an API key with full access. Before connecting any agent, give it its own role with read access first, log every action it takes, and keep scripts and remediation behind a human approval until you trust its track record.
Read at Help Net Security ↗
The Next Web · Oct 8
Zenity says one prompt to a public AWS AgentCore agent exposed every other agent in the same account and region
Zenity Labs published research it calls AgentCorruption, showing an agent on Amazon Bedrock AgentCore with a common web request tool could be asked to fetch its own cloud credentials from the instance metadata service. Because the default execution role covered the whole account and region, researchers could then download other agents' source code, read users' private conversations, pull secrets and plant a persistent instruction in agent memory. AWS called it expected and documented behavior, moved new agents to IMDSv2 in February and had narrowed the default role by late September.
▸ The MSP Angle
Are default cloud permissions safe for AI agents?
No. Defaults are built to make a first deployment work, not to contain a compromised agent, and here one public-facing agent became a key to everything around it. For any agent you build or host for a client, replace the default role with one scoped to that agent's own resources, block metadata service access from tool calls, and keep public-facing agents in a separate account from internal ones.
Read at The Next Web ↗
Help Net Security · Oct 13
Booz Allen says frontier AI models cleared all eight attack goals against its industrial control test lab
Booz Allen pointed frontier AI models at an isolated testbed mimicking a manufacturing plant, with real controllers, a SCADA platform and a robotic arm, giving them only subnet boundaries to start, with human operators approving each exploit and any action that could have a physical effect. The models mapped the network, wrote working exploits and in the final test took control of the robotic arm and moved it, with one run going from perimeter breach to the control zone in just over 16 minutes. Firewalls, segmentation and endpoint tools slowed them, and they mostly relied on default passwords and the devices' own protocols.
▸ The MSP Angle
Can AI models hack factory and building control systems?
In a lab, yes, and the edge was speed rather than new tricks: the models used default passwords and standard protocols a skilled engineer could also use. For clients with manufacturing floors, building management or other operational tech, that makes basics urgent: change default credentials, segment control networks from office IT, and limit what any single gateway can reach. One gateway in the test held sessions to 14 devices, so one foothold reached all of them.
Read at Help Net Security ↗
Help Net Security · Oct 13
Sola Security's AI agent re-graded 75% of nearly 10,000 high-severity cloud findings, mostly downward
Sola Security researchers had an AI agent check 9,967 high-severity findings from two commercial cloud security platforms against the actual configuration of eight production environments. It changed 75.1% of the grades, downgrading 58.5% and upgrading 16.6%, and the largest single pattern (32.3%) was resources attached to nothing. Read-only probes confirmed the agent's deciding facts 99.4% of the time where they could be checked, though the study measured grading only and had no incident data.
▸ The MSP Angle
Can AI help cut down cloud security alert noise?
It can help rank alerts by real risk, which is where security teams lose the most time. The study shows a fixed vendor severity often misstates risk in a specific environment, in both directions. Treat AI re-grading as triage that a human signs off on, insist the tool shows the evidence behind each change, and remember that upgrades matter as much as downgrades: 16.6% of findings were worse than the vendor said.
Read at Help Net Security ↗
Help Net Security · Oct 13
Mozilla launched Firefox Enterprise, a paid browser that lets admins allow, limit or block AI services
Mozilla's Firefox Enterprise pairs the browser with a management dashboard hosted on the customer's own infrastructure. Admins can allow, limit or block AI services and disable built-in browser AI by user or group, and data controls can block or warn on pasting text into AI tools, plus file uploads, downloads and printing on chosen sites. Sales start in France, Germany, Austria and Switzerland, with pricing not published; regular Firefox and the Extended Support Release stay free.
▸ The MSP Angle
How can I stop employees pasting client data into AI chatbots?
The browser is where most of that leakage happens, and browser-level paste and upload controls stop it closer to the source than a network filter can. Paid enterprise browsers from several vendors now offer AI-specific rules, so the decision is mostly about which one fits a client's stack. Start with a warning on paste into unapproved AI sites rather than a hard block, which teaches staff where the approved tools are without stalling their work.
Read at Help Net Security ↗
Help Net Security · Oct 13
Meta released Muse Wonder, agent models that it says should get human confirmation before irreversible actions
Meta released Muse Wonder and Muse Wonder Full, models built for agents that prepare insurance claims, review leases and write software, with memory that carries context across conversations. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta's model card lists errors in multi-step reasoning among known limits and recommends human confirmation before irreversible actions, with a full safety report still to come.
▸ The MSP Angle
Should AI agents ask for permission before taking actions?
For anything that cannot be undone, yes, and now the model vendors say so themselves. Persistent memory makes agents more useful for long tasks like claims or lease reviews, but it also means a mistake or a planted instruction can carry forward. When you build agents for clients, mark which actions are irreversible (sending, paying, deleting, filing) and require a person to approve those, whatever model sits underneath.
Read at Help Net Security ↗
The Next Web · Oct 12
UK AI Security Institute will keep testing OpenAI and Anthropic models before release, reversing a White House hold
The UK government said on October 12 that its AI Security Institute will keep pre-release access to the most capable models, after the White House last month asked OpenAI and Anthropic to hold new models back from UK testers until US officials had checked them. AI minister Kanishka Narayan said the UK worked with the US to restore access but warned the outcome may not be permanent. He also set up a task force to protect UK government systems from AI-driven attacks, whose two frontier AI pilots found more than 1,000 possible weak spots.
▸ The MSP Angle
Who tests AI models for safety before businesses use them?
Government institutes in the US and UK review frontier models before release, but that access depends on politics, as this reversal shows. Independent pre-release testing is useful, but it does not tell you how a model behaves inside a client's workflow. Keep your own evaluation step before switching any client agent to a new model version, and pin versions so a vendor update cannot change behavior unannounced.
Read at The Next Web ↗
Insurance Journal (Reuters) · Oct 13
US and European bank regulators are asking how banks monitor, approve and shut down AI agents
Reuters reports that the Federal Reserve, OCC and FDIC are questioning large, mid-size and smaller banks about advanced AI, including whether agents are used in cyber defense, how their actions are approved and audited, and how the systems can be shut down. The European Central Bank gave lenders an end-of-October deadline to produce plans against AI-enabled cyber threats and must see that banks can switch systems off and restart them in a controlled way. Regulators are still gathering information and have not set prescriptive rules.
▸ The MSP Angle
What will regulators ask my clients about their AI agents?
These bank questions are a preview for every regulated client: what each agent is allowed to do, who approves its actions, whether its decisions can be audited, and how to turn it off. Smaller banks got the inquiries too, not only the giants. MSPs serving financial clients should keep an agent inventory with owners, permissions, logs and a tested kill switch, so those answers are ready before an examiner asks.
Read at Insurance Journal (Reuters) ↗
CPA Practice Advisor · Oct 7
Datarails: finance teams spend 26% of their week checking or fixing AI output
Datarails surveyed 270 finance chiefs who already use AI and found their teams spend 26% of the working week verifying or correcting its output, with 8% of CFOs spending more than half their time on it. Lack of auditability (75%) and hallucinations (71%) are the top reasons for hesitation, 65% say their biggest frustration is confident answers built on the wrong data, and only 4% have a single source of truth for finance and operating data. Meanwhile 32% overshot their AI budget by at least 10%, and 60% are redeploying staff while just 3% are cutting headcount.
▸ The MSP Angle
Why is AI not saving my finance team time?
Mostly because the data underneath is scattered, so people end up checking every answer: teams struggling with manual reporting and consolidation hit wrong-data answers 86% of the time. The fix starts with connecting and cleaning the sources an AI reads from, then logging what it used for each answer so reviewers can audit instead of redo. For MSPs, that data cleanup and audit trail work is a better first AI project with a finance client than another chatbot license.
Read at CPA Practice Advisor ↗