SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

How to red team MCP agents by attacking their tools before attackers do

0
·14 views

Red teaming MCP agents means testing your own agent through its tools to find weaknesses before adversaries do. Most MCP security work covers the server layer, such as OAuth, tokens and transport, but misses the model layer, where agents may follow hidden instructions in tool descriptions, outputs or altered tools. The article describes threats including tool poisoning, rug pulls, tool shadowing and poisoned tool outputs, citing Invariant Labs experiments that showed these attacks in practice. It recommends a staged process: mapping tools and recording baselines, planting poisoned test tools, injecting instructions into outputs, and scoring results with automated pass/fail policies in CI/CD. Testing should run in staging with fake data and test accounts rather than production credentials.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Owners raise reliability concerns twice as often as reviewers, study finds

A research analysis published on bettascore.ai compares how frequently product reviewers and actual owners mention reliability. According to the piece, owners bring up reliability roughly twice as often as reviewers do. The finding suggests that long-term ownership experiences surface durability issues that professional reviews may not capture. The article's source text provides only the headline and link, so the specific products, sample size, and methodology are not available here.

0 comments Read more at Hacker News
0
ProgrammingDEV Community ·

Sites load about three times more high-risk trackers for US visitors than for EU visitors

A test of 200 popular websites, run on the same day from servers in Virginia and Frankfurt, found that US visitors triggered far more high-risk tracking requests and third-party cookies before any consent. On average, the Virginia visit saw about 12.6 high-risk tracking requests against 3.8 from Frankfurt, and 13.2 third-party cookies against 3.1. Cookie banners also varied by location, with 21 sites showing a banner only to the Frankfurt visitor. The difference is likely due to consent tools that adjust behaviour based on the visitor's country, since EU and UK rules generally require opt-in while most US rules are opt-out. The findings suggest that website teams testing from a single location may miss what visitors in other markets actually experience.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

DreamHost Forces Password Resets for Email and FTP Users in Ongoing Incident

DreamHost required email and FTP/SFTP users to reset their passwords after an incident that began with connectivity problems in its File Manager interface on October 23. The company blocked access for all users until resets were completed, calling the step a precaution, and sent recovery instructions in batches over the weekend. As of reporting, the incident remained active and File Manager was still unavailable. DreamHost advised account holders to check their mail forwarding rules and remove any unrecognised destinations, but said it cannot directly contact individual mailbox or FTP users. BleepingComputer reported that DreamHost had not answered its questions about the number of affected accounts or whether files or messages were exposed.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Anthropic Partners With Six Drug Developers to Use Claude in Rare Muscle Disease Research

Anthropic is working with six medicine developers, including Bristol Myers Squibb, to apply its Claude AI to finding treatments for a rare childhood muscle-wasting disease. The company has launched Anthropic Drug Discovery programs targeting rare childhood muscular dystrophy. Other named partners include BigHat Biosciences, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. Anthropic says Claude handles much of the scientific and operational work in these internal programs, though it has not disclosed commercial terms or whether any Claude-assisted work has entered formal clinical testing.

0 comments Read more at DEV Community