How to avoid escape sequence attacks in terminals?

VT100 terminals (which all modern terminal emulators emulate to some extent) supported a number of problematic commands, but modern emulators or distributions disable the more problematic and less useful ones. Here's a non-exhaustive list of potentially risky escape sequences (not including the ones that merely make the display unreadable in some way):

  • The arbitrary log file commands in rxvt and Eterm reported by H.D. Moore. These are indeed major bugs, fortunately long fixed.
  • The answerback command, also known as Return Terminal Status, invoked by ENQ (Ctrl+E). This inserts text into the terminal as if the user had typed it. However, this text is not under control of the attacker: it is the terminal's own name, typically something like xterm or screen. On my system (Debian squeeze), xterm returns the empty string by default (this is controlled by the answerbackString resource).
  • The Send Device Attributes commands, ESC [ c and friends. The terminal responds with ESC [ … c (where the can contain digits and ASCII punctuation marks only). This is a way of querying some terminal capabilities, mostly obsolete but perhaps used by old applications. Again, the terminal's response is indistinguishable from user input, but it is not under control of the attacker. The control sequence might look like a function key, but only if the user has an unusual configuration (none of the usual settings I've encountered has a valid function key escape sequence that's a prefix of the terminal response).
  • The various device control functions (DCS escapes, beginning with ESC P).
    • I don't know what harm can be done through DECUDK (set user-defined keys) on a typical terminal emulator.
    • DECRQSS (Request Status String) is yet another command to which the terminal responds with an escape sequence, this time beginning with \eP; this can be problematic since \eP is a valid key (Alt+Shift+P).
    • Xterm has two more experimental features: ESC P + p … and ESC P + q …, to get and set termcap strings. From the description, this might be used at least to modify the effect of function keys.
  • Several status report commands: ESC [ … n (Device Status Report). The terminal responds with an escape sequence. Most of these escape sequences don't correspond to function key escape sequences. One looks problematic: the report to ESC [ 6 n is of the form ESC [ x ; y R where x and y are digit sequences, and this could look like F3 with some modifiers.
  • Window manipulation commands ESC [ … t.
    • Some of these allow the xterm window to be resized, iconified, etc., which is disruptive.
    • Some of these cause the terminal to respond with either a escape sequence. Most of these escape sequences look low-risk, however there are two dangerous commands: the answers to ESC [ 2 0 t and ESC [ 2 1 t include the terminal window's icon label and title respectively, and the attacker can choose these.
    • At least under Debian squeeze, xterm ignores these commands by default; they can be enabled by setting the allowWindowOps resource, or selectively through the disallowedWindowOps resource. Gnome-terminal under Ubuntu 10.04 implements even the title answerbacks by default. I haven't checked other terminals or versions.
  • Commands to set the terminal title or icon name. Under xterm and most other X terminals, they are ESC ] digit ; title ESC \. Under Screen, the escape sequence is ESC k title ESC \. I find the concern over these commands overrated. While they do allow some amount of mischief, any web page has the same issue. Acting on a window based solely on its title and not on its class is akin to opening a file whose name was given to you by an untrusted party, or not quoting a variable expansion in a shell script, or patting a rabid dog on the nose — don't complain if you get bitten.

I find Varnish's response disingenuous. It's feels like it's either trying to shift the blame, or in security nazi mode (any security concern, genuine or not, justifies blackballing a feature).

The wisdom of terminal-response-escapes in general have been questioned at regular intervals, but still none of the major terminal emulation programs have seen fit to discard these sequences, probably in a misguided attempt at compatibility with no longer used 1970'es technology. (…)
Instead of blaming any and all programs which writes logfiles, it would be much more productive, from a security point of view, to get the terminal emulation programs to stop doing stupid things, and thus fix this and other security problems once and for all.

Many of the answerbacks are useful features: an application does need to know things like the cursor position and the window size. Setting the window title is also very useful. It would be possible to rely entirely on ioctl calls for these, however this would have required additional code and utilities to make these ioctl calls and transate them into unix-style text passing on file descriptors. Changing these interfaces now would be a lot of work, for little benefit.

Text files are not supposed to contain non-printing characters such as control characters. Log files are generally expected to be text files. Log files should not contain control characters.

If you're worried that a file might contain escape sequences, open it in an editor, or view it with less without the -r or -R option, or view it through cat -v.

It's not quite that simple; lots of people have code to set the xterm title as part of the prompt or etc., and for very good reasons (as anyone who's shutdown the wrong machine from the wrong terminal window can tell you!). Properly fixing it therefore involves introducing security contexts and thereby seriously complicating the interaction between shells and terminal emulators and probably people's dotfiles as well. Or you can go with the low rent solution of sanitizing anything that might be displayed in a terminal; this largely reduces to escaping control characters, which should probably be done anyway just to make them stand out (since in a log file they can indicate someone trying to inject shellcode).

(As an aside, punycode is a more severe instance of the same kind of problem — and has nevertheless been made an official standard. Sometimes security comes down to mitigating insecure designs that are outside one's control.)