Adversa AI launches a security platform to block attacks on coding agents in real time

Adversa AI has launched a coding-agent security platform, with the central claim that it blocks cyberattacks against AI agents in real time, according to a PR Newswire release dated October 29. The announcement is compact but pins down three things at once: the category is security, the protected asset is the coding agent, and the selling point is speed.

Picking coding agents as the beachhead makes sense. These agents carry far more privilege than a chat assistant — they read and edit code, run commands, call external tools, and often hold repository and CI credentials. Trick one into following a malicious instruction and the consequence is not a bad paragraph of text but a real commit and a leaked secret.

The entry points multiply accordingly. Issue and pull-request descriptions, dependency READMEs, log output and web documentation all get pulled into the agent's context, and any of them can carry instructions aimed at the model. A development workflow feeds outside content into an agent more densely than most business processes do.

The word that needs testing is real time. Security controls on an agent pipeline usually sit in one of three places: filtering inputs before the call, intercepting actions during execution, or auditing and rolling back afterwards. The cost, false-positive rate and strength of protection differ sharply between them, and a real-time blocking claim can refer to any of the three.

Buyers should ask concrete questions. How does it tell a legitimate tool call from a link in an attack chain? Which agent frameworks and runtimes does it support? Is deployment an inline gateway or a sidecar proxy? And what does it cost when a false positive blocks a legitimate build? Those answers decide whether the thing can stay switched on in production.

The wider context is that enterprises are issuing identities and permissions to agents, and non-human identities are starting to rival or outnumber employee accounts. Identity and access management, long organised around people, is being rewritten around agent credentials, scopes and behaviour auditing. Security vendors moving into that gap is an obvious step.

Three things to watch: whether Adversa AI publishes reproducible attack samples and interception-rate testing; whether the platform integrates with mainstream code hosting and CI tools; and whether rivals turn agent runtime protection into a category of its own. If they do, this layer may follow the path of EDR and move from optional to default.