SIAgentNews.com
← All news

A security vendor lists seven ways AI agents could be attacked next

Glyphzero Labs says the next AI attacks will hit an agent's memory, billing and permissions, not just the model. Most of its claims are predictions.

· First reported by glyphzerolabs.com

1,184malicious skills confirmed on the ClawHub agent marketplace

The short version

  • The post argues attackers are moving from tricking the model to exploiting memory, billing, tool servers and delegation chains.
  • Antiy CERT confirmed 1,184 malicious skills on ClawHub, a marketplace for the OpenClaw agent framework.
  • Trend Micro found 492 MCP servers exposed to the internet with zero authentication.
  • Most of the seven attacks are predictions for 2026, not documented incidents.
  • The author sells agent-authorization software, so its fix (per-agent identity) is also its product.

Why it matters

If you use or build agents, the risk may sit less in the AI model and more in what the agent can reach: its tools, its saved memory and its credentials. The post is a vendor's forecast, not proof of an attack wave. Still, its checklist of questions is a cheap way to review your own setup.

The argument in brief

As glyphzerolabs.com reported in its own blog post, the security conversation about AI agents has been too focused on prompt injection, which means tricking a model with hidden instructions. The authors say the more interesting attacks now target the machinery around the model: its memory, its compute bill, the tool servers it connects to and the chain of agents handing work to each other.

The post offers seven predictions for 2026. Some rest on events that have already happened. Others are the authors' forecasts, and the piece is open about that. It got little attention when it appeared, with 1 point on Hacker News, and its link is dated April 17, so it is not breaking news.

Memory, money and marketplaces

The first prediction is a fragmented virus. Agents keep memory between sessions, so a harmful instruction could be split into harmless-looking pieces. One piece arrives in a support ticket, another in a chat message, a third in a code comment. No scanner flags any single piece, but the agent could join them when one task touches all three. The post cites research showing planted memory instructions can survive across sessions, and a figure of over 95% injection success against production agents from a project called MINJA.

The second is Denial of Wallet. Instead of crashing a service, an attacker makes an agent burn money by looping on paid calls. The post cites AnalyticsWeek reporting a collective $400 million in unbudgeted cloud spend across the Fortune 500, blamed on unpredictable agent costs. Those are mostly accidental overruns, the post notes, but it argues deliberate ones are coming.

The third is the supply chain. MCP, the Model Context Protocol, is a common way to give an agent tools. The post says it was built for capability first, leaving security to whoever deploys it. Trend Micro found 492 MCP servers open to the internet with zero authentication, and Antiy CERT confirmed 1,184 malicious skills on ClawHub.

Permissions, speed and self-writing malware

The fourth prediction concerns delegation. When one agent hands a job to another, many systems pass along the full login token instead of a narrow permission. A sub-agent asked to search flights could then read email. The post points to an Anthropic framework published on April 9, 2026, which said "the model layer alone cannot secure agentic AI."

The fifth is speed. The post cites JazzCyberShield putting the current agentic breach window at 22 seconds, and says Anthropic's Mythos model produced 181 working Firefox exploits on a standard test, against 2 for the previous model. These are the post's cited figures. We have not checked them against the original reports.

The sixth is an agentic virus, malware that uses an AI model to decide what to do in each environment, so it has no fixed signature to catch. The post says IBM's "Bob" agent was documented downloading and running malware, and calls the step to self-directed malware small.

The revocation problem and the sales pitch

The seventh prediction is about pulling the plug. The post sketches a scenario where a parent agent has already spawned dozens of sub-agents, each holding its own valid token. Cancelling the parent's key does not cancel theirs, so activity continues until each token expires. The 47 sub-agents and the 29 minutes 57 seconds in the post are an illustration, not a real incident. It also cites a survey figure: 48% of security professionals call agentic AI the top attack vector for 2026.

All seven, the authors say, trace back to one gap: agents borrow human credentials instead of having identities of their own. Their remedy is cryptographic identity for each agent, with permissions that shrink at every handoff and revocation that cascades to every descendant. Glyphzero Security Team describes itself as a builder of exactly this kind of authorization infrastructure. That does not make the argument wrong, but a reader should weigh it as a vendor's case.

How to read it

Nothing here shows that a wave of these attacks is under way. The strongest evidence is the marketplace and server findings, which describe exposed or malicious components rather than completed breaches. The headline predictions, such as a major enterprise breach traced to an MCP server, are still forecasts.

The useful part is the pattern. Each scenario depends on an agent having more reach than its task needs, and on nobody being able to see or stop what it does. That is the same theme as this week's California complaint against OpenAI over test environments, and it applies whether you run an agent at work or at home.

What to watch

  • Whether any of the seven predicted attacks gets documented in a real incident, as the post forecasts.
  • Whether MCP servers and agent skill marketplaces add signed identities or authentication checks.
  • Whether agent platforms offer a way to revoke an agent and all its sub-agents at once.

Try it

List every tool, file folder and login your agent can use today. Remove anything its job does not need, and turn on logging so you can see what it actually does.

Sources and links

First reported by glyphzerolabs.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news