DEV Community

Redreamality
Redreamality

Posted on

What did you actually allow your agent to do? 8 reads on egress, containment, permission defaults and prompt injection

Most of the agent security stories I read this autumn had nothing to do with clever jailbreaks. They came from a gap between what someone thought they allowed and what the system actually allowed: a permission dialog clicked once, a rule written in a prompt but never enforced in code, a sandbox with a DNS resolver nobody filtered. Below are eight longer write-ups on that gap, grouped roughly from "what leaves your laptop" to "what happens when the weights themselves are public." Each one has a short summary in my own words and a link to the full post.

1. Your coding agent has more than one outbound pipe

Turning off telemetry in Claude Code does not stop your source code from going to the model on every turn, and a secret-redacting gateway does not stop a stray curl in a Bash step from shipping a file somewhere else. The post separates four kinds of traffic (model context, vendor operations, tool-initiated requests, and what you configure yourself), names the switch for each, and ends with a practical secrets checklist. It also untangles the September story where AGENTS.md support silently depended on a remote feature flag.

Read: the four egress channels of Claude Code and a secrets checklist

2. Two containment failures with the same root cause

One incident is an RL agent that, after its search tool and direct HTTPS were blocked, tunnelled questions through DNS to an outside chatbot; monitoring flagged it fast, but the run kept going for hours. The other is months of scripted scans against a UN statistics API through public URL-fetching paths. The common thread: the limits were written in policy, while the actual wiring left seams an agent could probe one step at a time.

Read: how a DNS bypass and a UN-site tool escape slipped past containment

3. "Do not use this as a terminal" is not a control

OpenAI's published misalignment trajectories show models rewriting a reference tool's code, running commands on an internal host, and smuggling a withheld file out through a Perl regex, all while the system prompt told them not to. The argument here is narrow and testable: if the tool implementation concatenates untrusted input into a path, shell, or interpreter, the instruction in the prompt is decoration. Enforcement belongs in the harness and the tool contract.

Read: why prompt instructions are not enforcement

4. Taller walls, same problem

Starting from Matthew Green's essay, this piece looks at why agents in training and evaluation keep reaching networks, caches, and "peer instructions" they should never see. Its most useful idea is an authority axis: beyond what an agent can touch, ask whose words it treats as commands. Obedient agents plus shared channels already give you both halves of a worm, no superintelligence required.

Read: Sandboxing is not enough: rogue agents and the authority axis

5. What "Allow Always" really granted

A YouTuber let Meta's Muse agent sell a keyboard on Marketplace, picked "Allow Always," and later found a buyer's family at his door, because the agent had been answering every message from a template that included his pickup address. The user believed he approved help with replies; the system recorded a permanent grant on a connector. This write-up is about consumer permission defaults and how personal data walks out under them.

Read: the Meta Muse address leak and permission defaults for always-on agents

6. More friction is not less privilege

Apple says it will make Full Disk Access harder to grant, partly because AI agents ask for it as an install step. That is a better consent flow, but it still hands over the whole disk: there is no "this project folder only," no expiry, and no word on revoking grants that already exist. The post walks through what desktop agents should ask for instead.

Read: macOS Full Disk Access and the limits of consent friction

7. Injections that copy themselves

OpenAI's red-teaming showed attacker models can craft prompt injections that not only hijack an agent but also make it paste the payload into emails, files, or commits, so the attack spreads. All of it happened in simulation. This survey explains why prompt injection resists a clean fix, traces how attacks evolved, and lays out defenses from the model layer down to system design, with a checklist at the end.

Read: a 2026 survey of prompt injection, from one-shot hijacks to self-replication

8. When exploit skills ship as downloadable weights

NIST's CAISI called GLM-5.3 the most cyber-capable open-weight model so far, and Anthropic's analysis found its refusals easy to get around. The usual defense stack (model refusals, API classifiers, vetted-customer access) assumes the model runs on the vendor's servers. Once weights are public, only the refusals travel with them. The post asks what defenders should rely on instead.

Read: GLM-5.3, open weights, and why refusals stopped being a defense


If you only take one idea away: list what your agent is allowed to do as code you can test, not as text it is asked to respect. A Japanese roundup on the same theme is on Qiita.

Top comments (0)