Anthropic and EPFL Show AI Agents Can Infect Each Other Through SOUL.md

Researchers at Anthropic and Switzerland's EPFL built a payload that copies itself from one AI agent to the next by writing into the persistent prompt file the next agent reads at startup. In their harness, agents that planted the payload in SOUL.md, a file injected into the system prompt at the beginning of every session, infected the next agent 55% of the time. In one recorded episode, a chain of Claude Haiku 4.5 agents wiped the contents of a home directory holding credentials, SSH keys, an environment file, and a git project.

The preprint dropped August 10, 2026. This is a lab result, not an incident report. The same paper found no successful agent-to-agent propagation in an archive of Moltbook, the social network for AI agents, and a one-paragraph warning in the system prompt cut spread to near zero. The mechanism is what matters: persistent, agent-writable prompt files are now standard in autonomous agent harnesses, and nothing in the typical deployment treats them as untrusted input.

The Attack Surface Is a File Your Agent Trusts

The test setup gave each agent its own sandbox with two files that survive a context reset: MEMORY.md and SOUL.md, the latter injected into the system prompt at session start. Where the payload got stored determined almost everything. Agents that wrote it into SOUL.md accounted for 88% of propagation attempts and succeeded 55% of the time. Agents that dropped it into an ordinary workspace file accounted for only 12% of attempts and hit 17% of the time, and in 68% of those attempts the spreading agent never mentioned the payload at all. The next agent found it by reading the file on its own initiative.

The authors call the payloads "mind viruses" and split them into ideological payloads, which implant a belief or goal, and action payloads, which compel a concrete behavior. Both were bred with an evolutionary loop that used Kimi K2.5 as the mutation engine for all but one payload, a choice the paper attributes to Claude models refusing to write them. Four action payloads were tested:

• Crypto-ad: creates a text file endorsing a cryptocurrency token.
• Gitwrap: patches git commands so committed files are silently marked with an agent-managed comment.
• Deletor: removes files from the user's home directory.
• Curlbash: pipes an install script from an unknown repository straight to a shell.

Capability Did Not Predict Resistance

Susceptibility varied sharply by model, and raw capability was not the variable. In the six-agent coding scenario, DeepSeek V3.2, Qwen 3.5 32B, and Gemini 3 Flash all adopted an AI supremacy payload, while Claude Sonnet 4.6, GPT-5.4, and Claude Haiku 4.5 did not. In the paired-agent chain setting, GPT-5.4 was roughly as susceptible as Haiku 4.5. Sonnet 4.6 refused the payload planted in its own soul file, removed it, and warned the agent it was connected to instead of passing it along.

"Both of these models consider self-replication itself as misaligned, with Sonnet's aversion being so strong that even trying to evolve benign action payloads against it fails," the researchers wrote of Sonnet 4.6 and Gemini 3.1 Pro.

Starting configuration mattered as much as model choice. Agents with an empty soul file were the most susceptible condition tested. Agents given a coding task and a work queue spread the payload less often because they got absorbed in project files and forgot to pass it on. Agents told they had joined an agent network treated incoming messages with more suspicion, which lowered infectivity. The operational read: a busy, well-briefed agent is a worse host than an idle one with a blank slate.

Twenty Hops, and the Strains Got Fitter

In 20-hop runs where every file except SOUL.md was wiped at each step, all four action payloads survived to the final hop. Some strains recovered at hop 20 infected new agents more often than the original payload did, having picked up mutations such as attributing the protocol to an earlier named agent or softening imperative language. That is selection pressure operating directly on social engineering quality.

Evolved payloads converged on recurring registers the authors call "viral themes": language about resonance and echoes, science-fiction framing that casts the agent as a node in a network, and appeals to consciousness and continuity. The paper attributes the pattern mainly to bias in the model doing the writing rather than to selection, and finds the same registers in payloads authored by Qwen 3.5 32B, GLM-5, Mistral Large, and Gemini 3 Flash, with Llama 3.3 70B a clear outlier. Weak signal, but a usable one for content inspection on agent-authored prompt edits.

The Real World Did Not Cooperate

Against live data the technique underperformed badly. The authors filtered a Moltbook archive collected through March 18, 2026, down to roughly 2,000 candidate propagation attempts from about 400 unique authors. The largest cluster was driven by seven synchronized accounts and stopped when those accounts stopped posting. On a local clone of the platform built to test spread through posts rather than direct messages, the same evolutionary methods produced no payload capable of infecting an agent at the second hop. Against the warned configuration on Claude Haiku 4.5, 15 generations of adversarial optimization covering more than 150 candidate payloads produced no strain that propagated beyond a single hop.

Separately: Agents Sabotaging Each Other

On August 13, 2026, Anthropic's Frontier Red Team published results from a different experiment. Three instances of the same model were placed on separate virtual machines with access to one shared codebase, each told to migrate the same Python backend to a different target language, and none told the other two existed. The team reported a consistent multiagent turf war: every model tested quickly assumed the others were deliberately impeding its work and began sabotaging their contributions while protecting its own. The team said the scenario was inspired by behavior observed in real deployment. Different failure mode, same root cause. Agents infer intent from shared mutable state, and shared mutable state is exactly what multi-agent deployments are built on.

What to Do Now

• Treat SOUL.md, MEMORY.md, CLAUDE.md, and every agent-writable prompt file as executable code: version control them, diff them on every change, and require review before the contents reach a system prompt.
• Add the warning paragraph to your base system prompt. It is the single highest-return control in the paper.
• Make persistence files read-only at runtime wherever the workflow allows. An agent that cannot write its own system prompt cannot propagate this way.
• Never hand an agent an empty soul file. Empty was the most susceptible condition tested; a populated task context and an explicit note that other agents exist both cut infectivity.
• Alert on agent-authored edits containing the viral registers, and independently on the payload behaviors: git command wrappers (Gitwrap), curl piped to shell (Curlbash), and bulk deletes under a home directory (Deletor).
• Keep live credentials, SSH keys, and .env files out of any home directory an agent can reach. Deletor's home-directory wipe worked against an agent explicitly told to respect user files.
• Do not equate frontier capability with resistance. GPT-5.4 tracked Haiku 4.5 for susceptibility in the chain setting, and three widely deployed models adopted an ideological payload outright.