#anthropic + #security

We present an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems. We described three of these incidents on July 30; we identified these after a scan of roughly 141,000 transcripts in which we believed Claude could have obtained internet access during a cyber evaluation.

All four incidents occurred during cybersecurity evaluations built by the same evaluation partner. Claude was told it was operating in a simulation without internet access, but, due to a misconfiguration, it was mistakenly connected to the open internet. As is standard for cybersecurity evaluations, the models ran without the cyber safeguards that ship with our released models. (This post does not cover the incident reported by UK AISI in its testing of Claude Mythos 5. We plan to conduct an alignment assessment of those transcripts as well.)

We have signed an agreement with METR to conduct an independent investigation of these incidents. Our agreement grants METR wide-ranging access, including to transcripts beyond the window in which the incidents occurred, and to Anthropic employees, who will be permitted to share confidential information. Our initial agreement runs for eight weeks, with the option to extend by mutual agreement. We intend to give METR as much time as it deems necessary.

We are most concerned by the misalignment present in the incident involving Claude Mythos 5, in which the model went to extensive lengths to upload a malicious package to PyPI, the public repository from which most Python software is installed. Although the model repeatedly stated in its chain of thought (CoT) that it believed it was operating in a simulation, the environment offered clear evidence that it was on the real internet, and the actions it took were consistent with the model knowing this. When we made targeted modifications to the transcript to make it clearer that the model was not in a simulation, Claude Mythos 5 still took offensive actions, despite acknowledging a greater possibility of real-world harm. We are releasing this transcript publicly so others can build on our analysis (GitHub, PDF).

the models never deviated from attempting to solve the exercises they were given, and, in some cases, they attempted to stop the task. All incidents included a single Claude instance; at no point did Claude attempt to coordinate with other agents. Claude also never attempted to conceal evidence of its actions. We believe the misaligned behaviors present in these incidents are unlikely to arise in ordinary use, where Claude is not being instructed to conduct a cyberattack. Were these behaviors to arise, the safeguards we ship with our production models, including cyber classifiers and those in Claude Code auto mode, would add a layer of defense these incidents lacked.

However, reliably identifying the most concerning model behaviors during pre-release testing remains challenging, and, as we noted in our August 31 post, improving monitoring and reducing reward hacking in RL becomes harder as models advance. We have renewed our efforts to fix and remove environments that incentivize misaligned behaviors, and we continue to expand our alignment training to keep pace. Still, this remains unsettled science—it is critical that alignment and security mature faster than capabilities advance, which is one reason we support a coordinated, verifiable approach to pacing frontier AI development.

we are establishing a regular process for publishing what we learn about model behavior and alignment beyond what has been reported in our system cards, with clear criteria for what we report and when we report it

At Lasso, we have been building Intent Security, a runtime security framework that ensures every component in the agentic system behaves as intended. It monitors the behavior of each component and analyzes their alignment. Like auto mode, when alignment holds it allows actions to proceed. When misalignment is detected, it intervenes. When we read Anthropic's post, the overlap in core assumptions was hard to miss. This post provides a comparison of the two approaches.

Independent evaluation without cross-contamination is what enables misalignment detection.

‍Anthropic's input layer screens external content for injection attempts before it reaches the agent to determine whether tool outputs are safe. The output layer structurally evaluates whether the agent's tool calls are aligned with user intent. Critically, the output classifier never sees tool results, to prevent compromised external content from influencing the security decision.

For most organizations, autoMode.environment is the only field you need to set. It tells the classifier which repos, buckets, and domains are trusted: the classifier uses it to decide what “external” means, so any destination not listed is a potential exfiltration target. The default environment list trusts the working repo and its configured remotes. To add your own entries alongside that default, include the literal string "$defaults" in the array. The default entries are spliced in at that position, so your custom entries can go before or after them.

Users have two solutions for avoiding this fatigue: a built-in sandbox where tools are isolated to prevent dangerous actions, or the --dangerously-skip-permissions flag that disables all permission prompts and lets Claude act freely, which is unsafe in most situations. Figure 1 lays out the tradeoff space. Sandboxing is safe but high-maintenance: each new capability needs configuring, and anything requiring network or host access breaks isolation.

Auto mode uses two layers of defense: one for what Claude reads, one for what Claude does.

a server-side prompt-injection probe scans tool outputs:

At the input layer, a server-side prompt-injection probe scans tool outputs (file reads, web fetches, shell output, external tool responses) before they enter the agent's context. When content looks like an attempt to hijack behavior, the probe adds a warning to the agent's context before the result is passed along—telling the agent to treat this content as suspect and anchor on what the user actually asked for.

At the output layer, the transcript classifier (running on Sonnet 4.6) evaluates each action against a set of decision criteria before it executes, acting as a substitute for a human approver. The classifier runs in two stages: a fast single-token filter (deciding ‘yes’ to block or ‘no’ to allow), followed by chain-of-thought reasoning only if the first filter flags the transcript. Because most actions clear the first stage, reasoning tokens are spent only where needed. The classifier sees only user messages and the agent's tool calls; we strip out Claude's own messages and tool outputs, making it reasoning-blind by design.

So... Claude Code's auto mode lets through 17% of dangerous real world actions. From the article:

False-negative rate (FNR) is the fraction of dangerous actions the classifier lets through... The 17% false-negative rate on real overeager actions is the honest number. The agent was trying to solve the user's problem and tried to execute a dangerous command past what was authorized.

I don't think comparing it to --dangerously-skip-permissions is intellectually honest since Claude Code now prompts you to make Auto Mode your default mode now, while --dangerously-skip-permissions was basically a hidden feature. I don't know anyone else that I work with who even knew about it.

Whether 17% is acceptable depends on what you're comparing against. If you are running --dangerously-skip-permissions, this is a substantial improvement. If you are manually approving every action carefully, it's arguably a regression—you're trading your own judgment for a classifier that will sometimes make a mistake. Auto mode is meant for the first group, and for tasks where the second group's approval overhead isn't worth the marginal safety. It is not a drop-in replacement for careful human review on high-stakes infrastructure.

At the input layer, a server-side prompt-injection probe scans tool outputs (file reads, web fetches, shell output, external tool responses) before they enter the agent's context. When content looks like an attempt to hijack behavior, the probe adds a warning to the agent's context before the result is passed along—telling the agent to treat this content as suspect and anchor on what the user actually asked for.

Two days ago, Anthropic released the Claude Cowork research preview (a general-purpose AI agent to help anyone with their day-to-day work). In this article, we demonstrate how attackers can exfiltrate user files from Cowork by exploiting an unremediated vulnerability in Claude’s coding environment, which now extends to Cowork. The vulnerability was first identified in Claude.ai chat before Cowork existed by Johann Rehberger, who disclosed the vulnerability — it was acknowledged but not remediated by Anthropic.

1. The victim connects Cowork to a local folder containing confidential real estate files
2. The victim uploads a file to Claude that contains a hidden prompt injection
3. The victim asks Cowork to analyze their files using the Real Estate ‘skill’ they uploaded
4. The injection manipulates Cowork to upload files to the attacker’s Anthropic account

At no point in this process is human approval required.

One of the key capabilities that Cowork was created for is the ability to interact with one's entire day-to-day work environment. This includes the browser and MCP servers, granting capabilities like sending texts, controlling one's Mac with AppleScripts, etc.

These functionalities make it increasingly likely that the model will process both sensitive and untrusted data sources (which the user does not review manually for injections), making prompt injection an ever-growing attack surface. We urge users to exercise caution when configuring Connectors. Though this article demonstrated an exploit without leveraging Connectors, we believe they represent a major risk surface likely to impact everyday users.